Replace faiss bfKnn with code that leverages our pairwise_distance api
and select_k api - by computing tiling over the inputs.

This lets us remove faiss as a dependency

This reverts commit 8eaba84.

Change didn't seem to build in CI

Also remove metrics processors - since is handled inside PW distance

This reverts commit fe728e9.

This is causing incorrect results, just use the faiss select_k call
instead

The faiss repo allows k values up to 2048, but we were limiting
to 512 or 1024 instead. This seems to be because of compiler
errors such as

```
ptxas error   : Entry function '_ZN4raft7spatial3knn6detail15select_k_kernelIdmLb1ELi2048ELi8ELi64EEEvPKT_PKT0_mmPS4_PS7_S4_S7_i' uses too much shared data (0x10000 bytes, 0xc000 max)
ptxas error   : Entry function '_ZN4raft7spatial3knn6detail15select_k_kernelIdmLb0ELi2048ELi8ELi64EEEvPKT_PKT0_mmPS4_PS7_S4_S7_i' uses too much shared data (0x10000 bytes, 0xc000 max)
ptxas error   : Entry function '_ZN4raft7spatial3knn6detail15select_k_kernelIdmLb1ELi1024ELi8ELi128EEEvPKT_PKT0_mmPS4_PS7_S4_S7_i' uses too much shared data (0x10000 bytes, 0xc000 max)
ptxas error   : Entry function '_ZN4raft7spatial3knn6detail15select_k_kernelIdmLb0ELi1024ELi8ELi128EEEvPKT_PKT0_mmPS4_PS7_S4_S7_i' uses too much shared data (0x10000 bytes, 0xc000 max)
```

when compiling with larger k values.

However, these only affect double precision with 64bit indices - and float32
works up to k=2048 (even with 64 bit indices).

The faiss repo allows k values up to 2048, but we were limiting
to 512 or 1024 instead. This seems to be because of compiler
errors such as

```
ptxas error   : Entry function '_ZN4raft7spatial3knn6detail15select_k_kernelIdmLb1ELi2048ELi8ELi64EEEvPKT_PKT0_mmPS4_PS7_S4_S7_i' uses too much shared data (0x10000 bytes, 0xc000 max)
ptxas error   : Entry function '_ZN4raft7spatial3knn6detail15select_k_kernelIdmLb0ELi2048ELi8ELi64EEEvPKT_PKT0_mmPS4_PS7_S4_S7_i' uses too much shared data (0x10000 bytes, 0xc000 max)
ptxas error   : Entry function '_ZN4raft7spatial3knn6detail15select_k_kernelIdmLb1ELi1024ELi8ELi128EEEvPKT_PKT0_mmPS4_PS7_S4_S7_i' uses too much shared data (0x10000 bytes, 0xc000 max)
ptxas error   : Entry function '_ZN4raft7spatial3knn6detail15select_k_kernelIdmLb0ELi1024ELi8ELi128EEEvPKT_PKT0_mmPS4_PS7_S4_S7_i' uses too much shared data (0x10000 bytes, 0xc000 max)
```

when compiling with larger k values.

However, these only affect double precision with 64bit indices - and float32
works up to k=2048 (even with 64 bit indices).

This is needed by the ANN benchmarks code, as well as cuml

cuml expects cosine distance to be (1-Cosine(a,b)), where the
pairwise_distances api was returning just Cosine(a,b). Revert

I'm seeing a bunch of test failures in cuml CI - and at least some of them
fail with L2Expanded distance and work with L2Unexpanded distance.

Only calculate norms once per input (rather than per-tile) when
calculating L2Expanded distance

For col-major inputs, we transpose the inputs since we only handle row-major
in the tiled_brute_force_knn call.  However, this was happening on the stream
from stream pool when given multiple partitions - causing dask errors later on.

Fix.