Umap with NumPy 2.0.0 fails due to return_inverse in unique

[seberg]

return_inverse returns something different even for axis=0. This was likely a mistake in NumPy 2, and we will probably revert it (not for axis=None). If you care for working correctly with NumPy 2.0 and have the time, adding:

inverse = inverse.reshape(-1)

would be safe way avoid a regression for this version either way.

xref https://github.com/numpy/numpy/issues/26738