perf: speed up flat tree traversal in modular decode

[hjanuschka]

This stores leaf predictors directly in flat tree nodes and uses an unrolled traversal macro in the hot prediction loop. It removes repeated enum conversion and reduces branching while keeping prediction behavior unchanged. Unsafe is used only for unchecked tree/property indexing where indices are validated during tree decode/flattening and guarded by invariants.

[github-actions]

Benchmark @ 05fe313

MULTI-FILE BENCHMARK RESULTS (4 files)
  CPU architecture: x86_64
  WARNING: System appears noisy: high system load (2.46). Results may be unreliable.
Statistics:
  Confidence:               99.0%
  Max relative error:        3.0%

Comparing: e883140e (Base) vs 7d77dfd6 (PR)

File	Base (MP/s)	PR (MP/s)	Δ%
bike.jxl	24.238	24.374	+0.56% ±2.7%
green_queen_modular_e3.jxl	7.867	7.811	-0.71% ±0.3%
green_queen_vardct_e3.jxl	23.852	23.817	-0.15% ±2.2%
sunset_logo.jxl	2.787	2.926	+5.02% ±1.3%

[veluca93]

Same comment at #714 and #715 :-)

[veluca93]

I modified the implementation not to use unsafe; the PR includes #787.

[jonsneyers]

Could you rebase now that #787 is merged, and do a benchmark to see the isolated effect of the tree traversal speedup compared to that new baseline?

[veluca93]

Performance Summary (Commit 1c05b2b)

Machine	Threading	Base MP/s	PR MP/s	Avg Improvement
desktop	Single	92.09	91.92	+1.70% ± 0.32%
desktop	Multi	91.94	91.99	+1.78% ± 0.39%
framework-desktop	Single	109.22	109.54	+1.51% ± 0.33%
framework-desktop	Multi	110.02	109.76	+1.50% ± 0.23%
pixel7a	Single (Fast)	32.59	32.71	+1.11% ± 0.38%
pixel7a	Single (Mid)	23.61	23.61	+1.18% ± 0.32%
pixel7a	Multi	33.12	33.15	+1.01% ± 0.39%

Detailed per-image results