Add fir_filter step and FIR highpass and bandpass designs

[galenlynch]

This PR adds a step that applies any FIR kernel, and separate functions that design the default kernels, as suggested on #4563:

• fir_filter(recording, taps) applies a FIR kernel by overlap-save, either one kernel for all channels or one per channel group. Its output doesn't depend on chunking.
• design_fir_highpass(freq_min, fs) and design_fir_bandpass(freq_min, freq_max, fs) design Kaiser-window kernels: −6 dB at the cutoffs, a 175 Hz transition at the low edge (and 2 kHz at the bandpass's high edge), and at least 45 dB down beyond. The default highpass has 463 taps and the default bandpass 503.

Best merged after #4853: without it, phase_shift followed by this filter runs an FFT of a slow length.

---

Speed. One thread, 1 s chunks of 384 channels, i9-12900K:

get_traces, int16 in and float32 out, best of 7; MS/s counts channel-samples. The IIR is sosfiltfilt, which computes in float64.

Saving 30 s through the filter keeps most of the gain and uses less RAM. Input and output are in RAM, so the time covers reading, filtering and writing but not the disk (median of 3; RAM is the peak above baseline):

	1 job	8 jobs
highpass, IIR	67 MS/s, 0.24 GB	273 MS/s, 2.19 GB
highpass, FIR	163 MS/s, 0.05 GB	548 MS/s, 0.88 GB
bandpass, IIR	60 MS/s, 0.24 GB	262 MS/s, 2.46 GB
bandpass, FIR	131 MS/s, 0.05 GB	488 MS/s, 0.88 GB

In a chain of phase_shift, the highpass and median common_reference, the FIR makes saving 1.15–1.17 times faster with main's phase_shift (given #4853), and 1.27–1.36 times faster with #4563's.

---

Response. The defaults aim to be no worse than the default Butterworth. On 383 channels of a Neuropixels 2.0 recording, compared with the probe's noise spec (69 nV/√Hz):

• Below 150 Hz the FIR passes more than the Butterworth, but less than 2.5% of the probe's own noise.
• Between 150 and 300 Hz it passes 13% less (highpass) and 5% less (bandpass).
• In the passband it is flat to 0.7%, where the Butterworth is up to 5.3% low (highpass, above 400 Hz) and 3.8% low (bandpass, 400 Hz to 4.5 kHz).
• Between 6 and 9 kHz the bandpass passes 30% less than the Butterworth.

Median spectrum over channels, relative to the probe's noise spec, for the 300 Hz highpass (top) and 300–6000 Hz bandpass (bottom).

Precision. A FIR has no poles, so float32 rounding errors don't build up, even at low cutoffs:

highpass cutoff	FIR in float32	sosfiltfilt in float32
300 Hz	5e-7	5e-6
10 Hz	3e-7	5e-3

rms error against a float64 reference, relative to the signal; 10 s of 32 channels.

Seams. Each output sample depends only on the len(taps) - 1 input samples around it, and the step reads exactly those, so chunked output equals whole-signal output up to float rounding. The PR adds check_chunk_invariance(recording, chunk_size, atol) to core.testing to check this for any preprocessor:

import numpy as np
from spikeinterface import NumpyRecording
from spikeinterface.core.testing import check_chunk_invariance
from spikeinterface.preprocessing import design_fir_highpass, fir_filter, highpass_filter

rec = NumpyRecording(np.fromfile("raw.dat", dtype=np.int16).reshape(-1, 384), 30_000.0)
iir = highpass_filter(rec, freq_min=300.0, dtype="float32")
fir = fir_filter(rec, design_fir_highpass(300.0, 30_000.0), dtype="float32")
for r in (iir, fir):
    print(check_chunk_invariance(r, chunk_size=30_000, atol=np.inf))  # largest seam error

Numbers

Largest seam error with 1 s chunks, on 10 s of a Neuropixels 2.0 probe (384 int16 channels, 2.34 µV per count, float32 output):

	IIR	FIR
highpass	0.0082 µV	0.00027 µV
bandpass	0.0093 µV	0.00019 µV

Shorter kernels miss the design goal. A 400 Hz transition (221 taps), for example, passes 2.2 times the Butterworth's power between 150 and 300 Hz.

Details

• Alignment. Output sample n is sum_j taps[j] * x[n + center - j]. The default center makes a symmetric kernel zero-phase, and center=0 makes the filter causal.
• Edges. The input is zero beyond the recording's edges, or reflected with add_reflect_padding=True, as in filter.
• Integer output is rounded.
• Offsets. offset_to_uV is scaled by the kernel's DC gain.
• Serialization. The taps are serialized, because for this step they are the parameters.
• Dependencies. Only scipy.fft and scipy.signal.

A follow-up PR adds method="fir" to highpass_filter and bandpass_filter, for users who don't want to design a kernel.

[galenlynch]

The test failures came from a module-level import scipy in test_fir_filter.py, which broke collection in the jobs without scipy; f795fbf moves it into the tests that use it.

"Load fixtures from 0.105.1" fails for a reason outside this PR. In .github/scripts/serialization/objects.py, the branch for versions after 0.105.0 expects set_probe to return a new recording:

else:
    rec_with_probe = rec.set_probe(probe)  # new API returns a new recording; portable across versions

In 0.105.1, set_probe sets the probe in place and returns None, so serialize_objects.py fails with 'NoneType' object has no attribute 'save'. 0.105.1 is the first release to take that branch. Something like this should fix it:

else:
    rec.set_probe(probe)
    rec_with_probe = rec

The job runs only for PRs that touch core, which this one does through core/testing.py.