perf: avoid extra pull_symbol call in unsigned RLE repeat path

[hjanuschka]

Carved out from the ongoing perf/autoresearch stream after PR #705 split work.

Change:

• in SymbolReader::read_unsigned_clustered_inline, after push_token consume repeat_count/last_sym directly instead of calling pull_symbol again.

Why:

• removes one extra call/branch in the hot repeat path
• behavior is unchanged

[github-actions]

Benchmark @ 7aa67b9

MULTI-FILE BENCHMARK RESULTS (8 files)
  CPU architecture: x86_64
  WARNING: System appears noisy: high system load (2.38). Results may be unreliable.
Statistics:
  Confidence:               99.0%
  Max relative error:        3.0%

Comparing: 462454f5 (Base) vs 4005ee63 (PR)

File	Base (MP/s)	PR (MP/s)	Δ%
bicycles.jxl	7.114	7.074	-0.56% ±3.0%
bike.jxl	25.155	25.265	+0.44% ±1.9%
delta_palette.jxl	6.164	6.133	-0.50% ±0.6%
green_queen_modular_e3.jxl	8.296	8.217	-0.96% ±0.8%
green_queen_vardct_e3.jxl	25.779	25.686	-0.36% ±1.9%
lz77_flower.jxl	3.321	3.325	+0.14% ±0.5%
patches_lossless.jxl	3.203	3.195	-0.25% ±1.5%
sunset_logo.jxl	2.681	2.692	+0.43% ±1.7%

[veluca93]

A more precise benchmark appears to be inconclusive. Converting this to a draft and leaving it for later evaluation.

[veluca93]

Closing this for now.