The bound can include (1 << window size), since the offset is shifted by
1. There is no point in encoding a 0-byte window, and increasing the
size for possible backreferences improves the correction ratio.

Reported by @timothyzillon, @sw17ch, @jibsen.

Closes #31.

Fix typo in CONTRIBUTING.md

API Changes:

1. The limits for the window and lookahead size options have changed.
The minimum lookahead bits option (-l) is now 3 (previously, 2), and the
maximum is the window bits - 1 (previously, equal to the window bits).

Bug Fixes:

1. There was an edge case in the internal buffering where window bits of
4 and lookahead bits of 2 could cause up to 3 bytes of data to be lost
from the end of the bytestream. Thanks to @unixdj for reporting this,
with detailed notes on how to reproduce it.

2. Using a lookahead bits setting equal to the window bits (e.g. -w 8 -l
8) could lead to an infinite loop. Aside from fixing the root cause, now
the lookahead size must be smaller than the window size -- an equal
lookahead size will always lead to worse compression, and should
probably have never been a valid setting. Found by @unixdj.

3. A few compiler warnings due to signed/unsigned warnings have been
resolved.

Other Improvements:

1. The logic for determining whether a pattern substitution is
worthwhile has been improved, potentially leading to better compression.
Decoders from earlier releases should have no problems processing the
output from newer encoders.

2. A benchmarking suite has been added, to measure the impact of changes
on speed and compression ratios.

3. The encoder and decoder state machines have been streamlined,
reducing overall complexity. Thanks again to @unixdj, who saw
opportunities to simplify things while investigating other issues, and
did most of the work.

4. Notes on example usage and configuration have been added to the README.

5. Initial notes for potential contributors (CONTRIBUTING.md) have been
added.

6. The theft-based property tests are now exploring the full state space
of the program. Previously, some of the test cases weren't varying the
window or lookahead sizes, only the input data. This led to a couple
bugs (1 and 2 above) going undetected. Now all three are changed, as
well as the input buffer size for the decoder.

Three cheers for actively maintained projects.

Conflicts:
	heatshrink_encoder.c

Merge several PRs and other changes on `develop` over in preparation for
0.4.0 release.

Closes #16.
(Addressed directly by 15ebaddc, and indirectly by several other changes.)

Closes #26.

Fix the bug where last 7 bits are not processed.

Test case:
$ echo -n aaaa | ./heatshrink -w4 -l2 | ./heatshrink -dw4 -l2 | hexdump -C

expected output:
00000000  61 61 61 61                                       |aaaa|
00000004

bug:
00000000  61                                                |a|
00000001

Conflicts:
	heatshrink_decoder.c

Download and cache the Canterbury Corpus, compress files in it with
various settings, and report on the compression ratios.

Closes #24.

This should stress the suspending/resuming as buffers fill more often.

bit_accumulator was not preserved between calls.

It's only called with count <= 8.

Delete some code that is no longer being used, but did not get deleted
during merging because of cherry-picking / out of order integration of
several concurrent pull requests.

Closes #22.

Conflicts:
	heatshrink_encoder.c

Apparently, gcc only warns about [u]int16_t comparisons on platforms
where sizeof(int) == sizeof(int16_t).  Seems like integer type promotion
happens here, and gcc only warns when it changes the comparison result.

The maximum buffer length is 1<<16 (when window size is 15), in which case
end takes values >=0x8000, and start between 0 and 0x8000, inclusive.
Which means we really want to compare unsigned pos to signed start
(except when start is 0x8000), so comparing them as either int16_t or
uint16_t would be incorrect.

But the difference between end-1 and start is always strictly less than
1<<widow_size, which makes it positive when interpreted as int16_t.
The only other case in which the difference between pos and start may
reach a positive value on 16-bit platforms is when indexing is used, the
window size is 15 and the backlog is empty, making start 0x8000, and pos
gets the sentinel end-of-list value 0xFFFF from the index.  This case is
handled by changing the sentinel value to 0 when the backlog is not full.
(Updated, see below.)

Conflicts:
	heatshrink_encoder.c

EDIT:
   Since commit 7e818d55 eliminates the backlog full calculations by
initially filling the backlog with zeroes (i.e., the backlog is always
full), the special case sentinel value of 0 is no longer necessary.

Closes #19.

Closes #18.

Using a lookahead size equal to the window size could previously lead to
an infinite loop (discovered and fixed by @unixdj, thanks). There really
isn't any benefit to having a lookahead size that large, as it only
applies when the entire input stream is the same. In all other cases, it
makes the compression ratio worse.

Closes #20.

BUG: When window size equals lookahead, the encoder enters an endless
loop after WINDOW_SIZE bytes are sunk into the buffer, unless at EOF.

Test:
$ echo -n 0123456789abcdefX | ./heatshrink -w 4 -l 4 >/dev/null

Closes #21

Closes #21

Fix signed/unsigned comparison

As reported by @unixdj, there is a case where a few bytes can be
dropped from the end of the bytestream when used with a window_sz2
of 4 and lookahead_sz2 of 2 (-w 4 -l 2):

    $ echo -n aaaa | ./heatshrink -e -w4 -l2 | ./heatshrink -d -w4 -l2
    a   # should be "aaaa"

While st_check_for_input can treat 7 bits as sufficient input when
-w is 4 and -l is 2, that creates a corresponding issue where 1
spillover bit from the previous byte leads to filler of 0b000 0000,
which is interpreted as a marker to repeat (0b0) from 1 byte back
(0b0000) for 1 byte (0b00), leading to a duplication of the last
byte of input. Using a w,l pair where w+l < 7 leads to trailing
bits that are ambiguous, so raise the minimum lookahead bits to 3.

This problem does not occur with -w 4 -l 3, or any other valid config.

This is to ensure the full state space is explored. malloc (and re-use,
after zeroing) large in-memory buffers for use in the tests, as these
will not potentially fill the way smaller stack-allocated ones can.