opt: allow bitCost == scaleLog in ZSTD_rescaleFreqs (fixes #4681)

[tvost2]

Fixes #4681

ZSTD_rescaleFreqs() initializes the optimal parser's symbol statistics from a
dictionary's entropy tables. For the litLength / matchLength / offCode streams it
reads a per-symbol cost with FSE_getMaxNbBits(), which by contract returns a
"fake cost" of tableLog + 1 for zero-frequency symbols (lib/common/fse.h).
LLFSELog / MLFSELog are 9 and scaleLog here is 10, so a dictionary
whose FSE NCount leaves a symbol in 0..MaxLL / 0..MaxML at frequency 0
makes bitCost == scaleLog, tripping

assert(bitCost < scaleLog);

in DEBUGLEVEL>=1 builds. Release builds are unaffected: the next line computes
1 << (scaleLog - bitCost) = 1 << 0 = 1, which is exactly the
/* minimum to calculate cost */ fallback, so the statistics are already correct
for this case.

The literal loop immediately above already handles the analogous Huffman
case with assert(bitCost <= scaleLog). This makes the three FSE asserts
consistent with it (< → <=); no behavior change beyond the assert.

Reproducer (from #4681, Aisle Research)

446-byte input to a dictionary_loader-style harness driving
ZSTD_createCDict_advanced with an optimal-parser strategy:

zstd_opt.c:195: ZSTD_rescaleFreqs: Assertion `bitCost < scaleLog' failed.
  #11 ZSTD_rescaleFreqs              lib/compress/zstd_opt.c:195
  #12 ZSTD_compressBlock_opt_generic lib/compress/zstd_opt.c:1115
  #13 ZSTD_compressBlock_opt2        lib/compress/zstd_opt.c:1451
  #14 ZSTD_buildSeqStore             lib/compress/zstd_compress.c:3444

Confirmed locally (DEBUGLEVEL=1): aborts before the change, [*] no crash
after.

Test

fuzzer.c gains test : optimal parser + dictionary with sparse FSE tables:
build a CDict with each optimal-parser strategy (btopt/btultra/btultra2)
from the 433-byte dictionary embedded in that reproducer (its FSE matchlength
NCount has zero-frequency symbols), then compress + round-trip. Fails at the
assert on dev, passes with this change. basicUnitTests green end-to-end.

Note

#4681 also suggests option (b): gate the FSE branch on
fse.matchlength_repeatMode == FSE_repeat_valid (etc.) and fall through to the
no-dict-stats path per stream when the dict NCount was not fully valid. That is
more principled but more invasive; happy to follow up with it if preferred. This
PR is the minimal correctness fix.

---

AI-assisted; reproduced, fixed, built and tested locally by me, and I take
responsibility for the change.