[FEA] pairwise distance primitives to output compressed format from scipy

Scipy’s pdist() [1] function outputs a compressed form of pairwise distances that, for valid metrics, eliminates the symmetry and redundant self-loop distances.

The compressed output array is a deterministic size with deterministic indexing so it should be straightforward to implement this right in raft’s pairwise distances.

For now, we can use the current pairwise distance outputs (called the squareform format) for pdist in cupy and convert to the compressed form. It’ll use extra memory temporarily, though, so once the primitive is able to output this directly then we can update cupy to use it.

cc @teju85 @mdoijade

[1] https://docs.scipy.org/doc/scipy/reference/generated/scipy.spatial.distance.pdist.html#scipy.spatial.distance.pdist

Start with the existing RAFT pairwise-distance primitive and compare its squareform output with scipy.spatial.distance.pdist indexing and size. Review how CuPy currently converts the output, then make the primitive produce the deterministic compressed representation directly and verify that the resulting distances and indexing match pdist.