AK: Search CircularBuffer matches from the end of memory

[timschumi]

This is particularly useful for compression, where we want to search through the lookback buffer for the smallest possible distance [towards the end].

Do note that the new AK::memmem_reverse has somewhat unexpected semantics around what an "offset" is, matching the very last character of the memory range with a needle of size 1 would yield the offset 1. Please check the test files for more examples of this. Would love to hear some opinions about that, but I can't figure out how to solve it any better way.

[nico]

Any observable changes in the filesizes produced by tar -c --lzma with this?

[timschumi]

Any observable changes in the filesizes produced by tar -c --lzma with this?

tar -c --lzma Documents Source Tests yields a bit-exact identical archive before and after this change (after about 30 minutes of compressing). I assume that the length restriction to 2 <= x < 3 for pure memmem hits makes this have a small enough impact that this case actually rarely happens in practice with more than one possible match.

[Hendiadyoin1]

2 <= x < 3

So 2?

[timschumi]

2 <= x < 3

So 2?

Sorry, yeah, it's effectively x = 2. I earlier had it noted down as x <= 3, but the x = 3 case is a validation fallback since the hash-based lookup can handle that already (and therefore it would get its preference from the hash lookup).

[timschumi]

Found a bunch of other issues in other places that may affect this PR, will draft until I fixed those.

[timschumi]

Applying the forward/reverse preference to the whole size range (instead of just the ones covered by non-hash search) indeed creates a noticeable size difference:

-rw-r--r--   1 anon    users    298.6 KiB   2026-01-06 01:29:04  ST_fwd.tar.lzma
-rw-r--r--   1 anon    users    281.2 KiB   2026-01-06 01:18:34  ST_rev.tar.lzma

Do note that in all cases (that I remember) we still pick the optimal match length-wise, so this size difference should be created purely by said distance encoding.

I'll clean up a few other issues beforehand separately and then open this PR back up for review.