Fix icoh normalisation term and add unit test

[tsbinns]

Two things in this PR:

• Adds a unit test for the coherence feature, including a sanity check that the connectivity values lie in an expected range using simulated connectivity data.
• Fixes a bug for icoh where it was not being normalised properly, causing connectivity values to not lie in the expect [-1, 1] range.

[tsbinns]

Denominator term should have the square root taken. Unfortunately the strategy for coh where you square the numerator and take the absolute value doesn't work for icoh since we extract only the imaginary part and this info isn't properly preserved.

[tsbinns]

Another note, I would argue it is beneficial to expose the nperseg parameter. With a high sampling frequency like 500 Hz, nperseg=128 is completely fine, but as sampling frequency decreases, this can come very close to the max possible. The result is coherence values which converge towards 1 (i.e. perfect correlation) regardless of the data passed in, and icoh values which are completely erratic.

At least if the end-user has control over nperseg, they can reduce this to reflect their sampling frequency. Doing so decreases the spectral resolution, but at least the connectivity values are still valid.

py_neuromodulation/py_neuromodulation/features/coherence.py

Lines 82 to 84 in 12d3ed9

	self.f, self.Pxx = welch(x, self.sfreq, self.window, nperseg=128)
	self.Pyy = welch(y, self.sfreq, self.window, nperseg=128)[1]
	self.Pxy = csd(x, y, self.sfreq, self.window, nperseg=128)[1]

What are people's thoughts on exposing this param as a setting?

[timonmerk]

Also @tsbinns, it definitely makes sense to expose nperseg. Tbh I did not even think about that, but especially for lower frequencies that will be necessary!