perf: FIR phase-shift alternative + apply_raised_cosine_taper extract

Adds a sinc-FIR alternative to the FFT-based PhaseShift path, and factors
the FFT-specific raised-cosine taper out of get_chunk_with_margin so
bounded-support FIR consumers don't have to pay for it.

Changes
-------
- PhaseShiftRecording(method="fft"|"fir", n_taps=32, output_dtype=None)
  - method="fir": 32-tap Kaiser-windowed sinc, numba-jit with prange over time
  - Per-channel kernels cached once per segment
  - Int16-native fast path: reads int16 directly, writes float32 when
    output_dtype=np.float32 (skips round-back-to-int16)
  - FIR margin = n_taps // 2 (16 for 32-tap default), vs FFT path's 40 ms
  - Default "fft" preserves existing behavior
- apply_raised_cosine_taper(data, margin, *, inplace=True) public function
  in spikeinterface.core.time_series_tools
- get_chunk_with_margin(window_on_margin=True) emits DeprecationWarning
  and delegates to apply_raised_cosine_taper; old behavior preserved

Whittaker-Shannon justification
-------------------------------
Both FFT and FIR paths are numerical realisations of the same ideal
sinc-interpolation implied by the sampling theorem.  FFT does it spectrally
(phase ramp = DFT of full sinc kernel); FIR does it in time against a
Kaiser-windowed 32-tap truncation of the same sinc.  Approximations: sinc
truncation (32 taps captures >99% of kernel energy for any d ∈ [0,1)) and
Kaiser windowing (-80 dB stopband, two orders of magnitude below NP's
~50 dB SNR).  Measured 0.19% spike-band RMS vs FFT on real NP 2.0 data.

Performance (1M × 384 int16 AIND pipeline, PS → HP → CMR, 24-core host)
-----------------------------------------------------------------------
                                 stock FFT   FIR         speedup
  single-call get_traces         82.1 s      6.28 s      13.1×
  same, f32 propagated           85.8 s      4.40 s      19.5×

  TimeSeriesChunkExecutor n=24   4.98 s      1.62 s      3.1× on top
                                                        of CRE n=24

Memory (same pipeline, n_jobs=24, chunk=1s):
  FFT peak RSS: 10.8 GB       FIR peak RSS: 3.9 GB (2.8× less)

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: galenlynch