perf: tighten VarDCT group coefficient hot loop

[hjanuschka]

This removes unnecessary VarDCT scratch/transform zeroing and tightens coefficient writeback indexing in the hot decode loop. It keeps decode behavior unchanged while reducing redundant memory work. Unsafe is used only for permutation/current-coefficient indexing with bounds guaranteed by validated coefficient order and loop limits.

[github-actions]

Benchmark @ c1e4804

MULTI-FILE BENCHMARK RESULTS (4 files)
  CPU architecture: x86_64
  WARNING: System appears noisy: high system load (2.27). Results may be unreliable.
Statistics:
  Confidence:               99.0%
  Max relative error:        3.0%

Comparing: e883140e (Base) vs e3b9da14 (PR)

File	Base (MP/s)	PR (MP/s)	Δ%
bike.jxl	24.474	24.345	-0.53% ±2.9%
green_queen_modular_e3.jxl	7.918	7.816	-1.28% ±1.1%
green_queen_vardct_e3.jxl	23.883	24.346	+1.94% ±1.0%
sunset_logo.jxl	2.784	2.624	-5.71% ±0.6%