Investigate changing the default `gamma` for SpectralClustering

The default of `1.0` often leads to 0s in the embedding arrays, leading to NaNs in the KMeans array.

Using `None`, which ends up using `1 / len(X)` doesn't seem to have this issue.

(will update w/ more info later).

Start by reproducing SpectralClustering with the default gamma of 1.0 and with gamma=None, then inspect the resulting embedding arrays and KMeans input for zeros and NaNs. Done means establishing whether the default should change and documenting the evidence and resulting behavior.

Written by the indexing model from the issue text.