FSE: fuse symbol spreading with table building on the fast path

[deadcaf3]

What and why

FSE_buildCTable_wksp() and ZSTD_buildFSETable() build their tables in two passes: spread the symbols into the table, then walk it in position order and read each symbol back. This PR fuses the two passes on the fast path (no low-probability symbols).

FSE_TABLESTEP() is odd for every valid tableLog, so the spread is an invertible permutation: position u holds the symbol at sorted index (u * inv) & (tableSize-1), with inv the inverse of the step modulo tableSize. Walking positions in order and reading symbols from the sorted layout removes the scattered writes and the read-back. FSE_invTableStep() (new, in fse.h) computes inv and asserts it.

Tables are bit-identical, so the format and the compressed output are unchanged. The low-probability path is left as is. No API change.

Performance

Default make flags. Apple M1 (clang 21), Intel Xeon 8481C and AMD EPYC 9B14 (gcc 13.3, Ubuntu 24.04).

Table build, ns per table, 256 random distributions with tableLog 5 to 9, min of 7 rounds, base and new interleaved:

CPU	FSE_buildCTable_wksp	ZSTD_buildFSETable
Apple M1	449 → 365 (−19%)	509 → 389 (−24%)
Intel 8481C	677 → 512 (−24%)	600 → 524 (−13%)
AMD 9B14	547 → 460 (−16%)	628 → 511 (−19%)

Decompression speed, zstd -b1 -B<chunk> -i3 on 3.4 MB of C source, two interleaved runs:

chunk	Intel	AMD	M1
4 KB	+0.9%	+2.1%	+1.6%
16 KB	+0.7%	+2.1%	+1.9%
32 KB	+0.1%	+1.1%	+0.8%
64 KB	−0.2%	+0.3%	0%

Compression speed within ±1%, compressed sizes identical. Not measured: other data sets, levels above 1, MSVC, 32-bit.

Verification

• New unit test in tests/fuzzer.c compares both builders against a plain spread-then-read-back reference for every tableLog, with and without low-probability symbols. It fails on three mutants, including a wrong but shared inverse that keeps encoder and decoder consistent.
• fuzzer, zstreamtest and decodecorpus -t pass with asserts enabled. make staticAnalyze reports the same findings on dev and on this branch.
• Applying the same fusion to FSE_buildDTable_internal() regressed 5 to 23% on Sapphire Rapids with gcc, so it is left out.

Checklist

• I searched open pull requests and issues: this problem is not already being addressed.
• Bug fix: it comes with a reproducer or a test that fails without the fix. (not a bug fix)
• Performance change: numbers from repeated runs are included, with CPU, compiler, flags and data.
• I stated below whether AI or other automated tools were used.