Distributed Neighborhood Attention S2

[azrael417]

This MR adds distributed Neighborhood Attention S2 support and fixes some issues in the existing attention kernel.

• existing serial attention kernel preallocated an output tensor which was too large when using attention based downsampling. This is fixed
• this kernel computes numerically the same for self-attention and cross-attention with downsampling.
• this MR adds distributed neighborhood attention along with some new tests for the feature. This kernel does not support up- or downsampling yet

[bonevbs]

Please bump the version number to 0.9.1a and start the Changelog for v0.9.1.

[bonevbs]

Please bump the version number to 0.9.1a1 and start the Changelog for v0.9.1

Also some minor comments

[bonevbs]

@rietmann-nv can you also have a look at the new CUDA kernels for distributed spherical attention

[mauro-bis]

Looks good to me.