perf: avoid zero-initializing modular channel payloads

[hjanuschka]

This adds an uninitialized image constructor that zeros only offset/padding regions, then uses it for modular channel allocation. It removes large buffer zeroing work in hot decode paths while preserving prediction boundary behavior. Unsafe is confined to allocation/row access internals and is safe because modular decode fully writes the data region before any read, and padding is explicitly zeroed.

[github-actions]

Benchmark @ 2c27f27

MULTI-FILE BENCHMARK RESULTS (4 files)
  CPU architecture: x86_64
  WARNING: System appears noisy: high system load (2.24). Results may be unreliable.
Statistics:
  Confidence:               99.0%
  Max relative error:        3.0%

Comparing: e883140e (Base) vs 5b261fb2 (PR)

File	Base (MP/s)	PR (MP/s)	Δ%
bike.jxl	24.177	24.113	-0.26% ±2.7%
green_queen_modular_e3.jxl	7.864	7.819	-0.57% ±1.6%
green_queen_vardct_e3.jxl	23.819	23.617	-0.85% ±1.4%
sunset_logo.jxl	2.780	2.796	+0.58% ±0.9%