Improve performance on generate_unit_locations

[h-mayorquin]

Currently the generate_unic_locations enforces the unit constrain with a simple rejection sampling algorithm that is non-deterministic:

spikeinterface/src/spikeinterface/core/generate.py

Lines 2007 to 2029 in 55c7de1

	if minimum_distance is not None:
	solution_found = False
	renew_inds = None
	for i in range(max_iteration):
	distances = np.linalg.norm(units_locations[:, np.newaxis] - units_locations[np.newaxis, :], axis=2)
	inds0, inds1 = np.nonzero(distances < minimum_distance)
	mask = inds0 != inds1
	inds0 = inds0[mask]
	inds1 = inds1[mask]
	if inds0.size > 0:
	if renew_inds is None:
	renew_inds = np.unique(inds0)
	else:
	# random only bad ones in the previous set
	renew_inds = renew_inds[np.isin(renew_inds, np.unique(inds0))]
	units_locations[:, 0][renew_inds] = rng.uniform(minimum_x, maximum_x, size=renew_inds.size)
	units_locations[:, 1][renew_inds] = rng.uniform(minimum_y, maximum_y, size=renew_inds.size)
	units_locations[:, 2][renew_inds] = rng.uniform(minimum_z, maximum_z, size=renew_inds.size)
	else:
	solution_found = True
	break

But I think this is a clear use case for the Bridson's algorithm for sampling on a poison disk with is log(N)

https://www.cs.ubc.ca/~rbridson/docs/bridson-siggraph07-poissondisk.pdf

• Time the current implementation for some channel locations
• Implement Bridson algorithm in Python
• Profile it.

This is a good first issue.