perf: optimize weighted predictor core and row updates

[hjanuschka]

This speeds up weighted prediction by removing derived-enum conversion overhead, unrolling hot error/weight paths, and reusing per-row predictor state. It also updates row initialization call sites so the optimized predictor state layout remains behavior-equivalent. Unsafe is used only for bounds-check elimination on indices that are guarded by row-size invariants and debug assertions.

[github-actions]

Benchmark @ 1d80d23

MULTI-FILE BENCHMARK RESULTS (4 files)
  CPU architecture: x86_64
  WARNING: System appears noisy: high system load (2.37). Results may be unreliable.
Statistics:
  Confidence:               99.0%
  Max relative error:        3.0%

Comparing: e883140e (Base) vs 3ca46430 (PR)

File	Base (MP/s)	PR (MP/s)	Δ%
bike.jxl	24.125	24.213	+0.37% ±1.9%
green_queen_modular_e3.jxl	7.889	7.263	-7.93% ±0.3%
green_queen_vardct_e3.jxl	23.835	23.684	-0.63% ±0.7%
sunset_logo.jxl	2.791	2.909	+4.25% ±0.9%