compile DynamicColoredNoise IIR loop with numba

[kylmcgr]

Summary

The per-sample Kasdin IIR recursion in DynamicColoredNoise was a pure-Python
double loop (over input bins × output samples) and the dominant cost in the
velocity→ecephys encoder. This ports the hot loop verbatim to a numba njit
kernel. The recursion is inherently sequential (output[i] feeds the delay
line that produces output[i+1]), so it can't vectorize over time — compiling
the scalar loop is the right tool.

No behavior change: the white-noise draw stays in NumPy (RNG sequence
preserved) and the MLX/array-api input path is untouched (the loop already ran
on a NumPy view). Channels are independent, so iterating over them inside the
sample loop is equivalent to the old vectorized-over-channels form; non-finite
β channels skip the coefficient update (an EMA-toward-self no-op).

Results

Per-stage encoder benchmark (bench_cosine_encoder.py, 256ch @ 30 kHz):

chunk	lfp.noise before	after	speedup
1 (live path)	0.686 ms/call	0.038 ms/call	~18×
50	33.85 ms/call	1.15 ms/call	~29×

End-to-end (simulator over LSL, 50 Hz, 256ch): PINK_NOISE dropped from
~0.70 ms/msg to 0.088 ms/msg — no longer the pipeline's longest stage.

Correctness

• Output matches the previous implementation to max abs diff 1.1e-13
  (floating-point reassociation only: the old code's np.einsum pole reduction
  sums in a different order than the scalar loop). Same-seed determinism holds.
• All 117 unit tests pass (tests/unit/). The compiled loop also cuts the
  suite's runtime from ~28.5 s to ~2.5 s.

Dependency

Declares numba explicitly (it was already present transitively via
ezmsg-event).