Currently, the angular distance can be computed in these methods using the inner product of the normalized vectors (or l2 of l2-normalized vectors) but this puts the burden of managing the norms and doing the normalization on the users. We should support cosine first-class right in the distance computation to alleviate these burdens.

Tagging @liliu-z for awareness and tracking.

• brute-force (https://github.com/rapidsai/raft/pull/1426/files)
• nn-descent & cagra ([FEA] CAGRA to support cosine and inner product distance  #1768)
• ivf-pq
• ivf-flat
• random ball cover