Optimize MLX simulator hot paths

[cboulay]

Summary

• compile and fuse the MLX cosine-encoding expression behind a lazy optional import
• keep large MLX signals on-device when adding line noise, with a measured size gate for small messages
• replace benchmark-style unit tests with state-aligned backend correctness tests
• add a standalone streaming benchmark that evaluates each MLX chunk immediately

Profiling

Measured on an M4 Pro with MLX 0.31.2:

• cosine encoder, 500 x 2 to 256: 690.8 us NumPy vs 177.0 us MLX, 3.90x speedup
• dynamic colored noise, 50 x 8 to 15000: 1038.9 us NumPy vs 1076.9 us MLX, effectively neutral
• line noise, 15000 x 256: 6078.4 us NumPy vs 476.4 us MLX, 12.76x speedup

The line-noise MLX path is gated at 200,000 signal elements because dispatch overhead loses below that crossover.

The colored-noise recurrence and baseline-drift hybrid prototypes were intentionally not moved to MLX: the real simulator shapes were faster with the existing Numba implementations.

Validation

• 142 tests passed
• ruff check passed
• ruff format check passed
• benchmark results validate NumPy and MLX outputs before reporting timings