compile SparseKernelInserter event loop with numba

[kylmcgr]

Summary

SparseKernelInserter._process looped over every event in Python and, per
event, re-evaluated the kernel (allocating a mask + zeros + fancy-index) and
did a sliced add. The cost scaled with channels × firing rate, making it a top
stage in the velocity→ecephys encoder.

This pre-samples every kernel once into a zero-padded table (one row per
distinct kernel, cached in state and rebuilt on stream reset) and scatters it
onto the dense buffer in a numba njit loop. Event value → table row is
resolved vectorized via np.unique over the (typically tiny) set of distinct
values.

The rewrite stays fully general — it works for any Kernel subclass
(ArrayKernel waveforms, FunctionalKernel, unit impulses) and preserves
scale_by_value, acausal pre_samples, and the chunk-boundary pending tail.

Results

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

chunk	kernel_insert before	after	speedup
1 (live path)	0.203 ms/call	0.039 ms/call	~5.2×
50	9.78 ms/call	1.68 ms/call	~5.8×

End-to-end (simulator over LSL, 50 Hz, 256ch): WAVEFORMS at 0.097 ms/msg,
flat-to-better despite each message now carrying 2× the samples.

Correctness

• Output is bit-identical to the previous implementation (max abs diff
  0.0) — kernel insertion is pure indexed addition, no reduction reassociation.
• 151 unit tests pass (was 147): adds 4 tests covering the new resolution
  paths — scale_by_value with a real kernel and with a MultiKernel, unknown
  values falling back to the default key, and an explicit default_key.

Dependency

Declares numba explicitly. It was already imported by poissonevents.py but
never declared — this fixes that latent gap.