[h-mayorquin]

This PR continues #4840 (Filter subsets of channels at a time to reduce peak memory), #4903 (Filter subsets of channels at a time in gaussian_filter to reduce peak memory) and #4904 (Shift subsets of channels at a time in phase_shift to reduce peak memory), now for common_reference. Once #4904 is in, it is the step that sets the peak memory of the usual Neuropixels chain on int16 traces.

Subtracting the float64 median from an int16 chunk makes a float64 copy of the whole chunk. Doing it in blocks of samples takes the peak from 5 to 1.2 times the chunk (6 to 2.6 times for the chain) and it is a bit faster. The output is identical to main except for the local average, where a few values can move by 1 depending on where the chunk is cut, as already happens on main. The local median is about 25% slower.

This touches the same lines as #4854 (Faster exact median for common_reference) by @galenlynch and I am fine with fixing the conflicts if that one goes first.

	main	this PR
global median, 1 s	501%, 0.086 s	117%, 0.071 s
global median, 10 s	501%, 0.853 s	102%, 0.704 s
local median, 1 s	301%, 0.520 s	110%, 0.643 s
chain, 1 s	601%, 0.525 s	259%, 0.514 s
chain, 10 s	601%, 6.707 s	255%, 6.573 s

Neuropixels 2.0 recording, 384 channels at 30 kHz, int16. Peak memory from tracemalloc as a percentage of the chunk (including the output) and time as the best of 3 runs. The chain is phase_shift from #4904, bandpass_filter and common_reference with the global median.