Improve phase shift memory efficiency

[h-mayorquin]

Been testing how much memory does phase shift uses.

Usings scipy instead of numpy I got from

18.3 GiB allocation to 10.3 GiB allocation.

Command line: /home/heberto/miniconda3/envs/neuroconv_env/bin/memray run --native --output /home/heberto/development/spikeinterface/build/memray-dev_profile.bin --force --follow-fork /home/heberto/development/spikeinterface/build/dev_profile.py
Start time: 2024-05-31 05:32:40.122000
End time: 2024-05-31 05:33:45.693000
Total number of allocations: 59354
Total number of frames seen: 2905
Peak memory usage: 18.3 GiB
Python allocator: pymalloc

to:

Command line: /home/heberto/miniconda3/envs/neuroconv_env/bin/memray run --native --output /home/heberto/development/spikeinterface/build/memray-dev.bin --force --follow-fork /home/heberto/development/spikeinterface/build/dev.py
Start time: 2024-05-31 05:30:53.499000
End time: 2024-05-31 05:32:06.583000
Total number of allocations: 64270
Total number of frames seen: 3711
Peak memory usage: 10.3 GiB
Python allocator: pymalloc

for the following script:

from spikeinterface.core import generate_recording_by_size
from spikeinterface.preprocessing import phase_shift
import numpy as np

recording = generate_recording_by_size(full_traces_size_GiB=1.0)
phase_shift_array = np.random.rand(recording.get_num_channels())
phase_shift_recording = phase_shift(recording, inter_sample_shift=phase_shift_array)
phase_shift_recording.get_traces()

I tested this against @alejoe91 data for the hackaton to see that the behavior is not changed and profiled the two (the chunk is medium size):

Following a suggestion by @DradeAW I also tested the speed differences in a smaller chunk:

My reading is that the timing differences are dwarfed by the memory improvements.

We can leave the keyword (probably better named use_scipy_fft) or not, probably better not to leak such details.