perf: tighten entropy and LZ77 hot paths

[hjanuschka]

This makes ANS/Huffman/LZ77 read paths more direct in hot loops by inlining and removing repeated bounds-check overhead where invariants are already validated. It also adds a no-LZ77 clustered read helper for specialization call sites. Unsafe is only used for unchecked indexing under validated context/window invariants, with debug assertions documenting each bound.

[github-actions]

Benchmark @ a1507cc

MULTI-FILE BENCHMARK RESULTS (4 files)
  CPU architecture: x86_64
  WARNING: System appears noisy: high system load (1.91). Results may be unreliable.
Statistics:
  Confidence:               99.0%
  Max relative error:        3.0%

Comparing: e883140e (Base) vs 5334cbe0 (PR)

File	Base (MP/s)	PR (MP/s)	Δ%
bike.jxl	24.486	24.392	-0.39% ±2.9%
green_queen_modular_e3.jxl	7.897	7.651	-3.12% ±0.5%
green_queen_vardct_e3.jxl	23.997	23.816	-0.76% ±1.1%
sunset_logo.jxl	2.784	2.977	+6.91% ±0.9%

[veluca93]

This PR clearly has some speed benefits, but I am not convinced by the approach to unsafe - I will look at this more in depth and I will try to see how it can be improved.