perf: optimize noise hot path and inline RNG helpers

[hjanuschka]

This speeds up noise rendering by removing per-element u64 splitting branches and reusing a direct u32 view of the RNG batch. It also inlines Noise::strength and Xorshift128Plus::fill in hot call paths. Unsafe is used only for a temporary [u64; N] to [u32; 2N] reinterpret cast, which is safe because the layouts are identical and the data is read-only.

[github-actions]

Benchmark @ 1a54821

MULTI-FILE BENCHMARK RESULTS (4 files)
  CPU architecture: x86_64
  WARNING: System appears noisy: high system load (2.35). Results may be unreliable.
Statistics:
  Confidence:               99.0%
  Max relative error:        3.0%

Comparing: e883140e (Base) vs bf26387b (PR)

File	Base (MP/s)	PR (MP/s)	Δ%
bike.jxl	24.244	24.178	-0.27% ±2.7%
green_queen_modular_e3.jxl	7.925	7.872	-0.66% ±0.9%
green_queen_vardct_e3.jxl	23.854	23.877	+0.10% ±1.4%
sunset_logo.jxl	2.788	2.786	-0.09% ±1.4%