[FEA] Lanczos solver v2

[lowener]

I unfortunately don't have permissions to push on @aamijar branch for the previous Lanczos solver PR (#2416) so I kept his commits and continued it here.

Lanczos Solver for Sparse Eigen Decomposition

We propose a new lanczos solver in raft that fixes the issues present in the previous solver raft::sparse::solver::detail::computeSmallestEigenvectors.

Specifically we address the following issues:

1. Numerical Stability for both float32 and float64 datatypes
2. Efficiency and Speed of Convergence

This new implementation is taken from the cupy library cupyx.scipy.sparse.linalg.eigsh where the thick-restart and full reorthogonalzation methods are used.

Additionally this PR exposes a python api for raft lanczos solver with an interface similar to scipy.sparse.linalg.eigsh and cupyx.scipy.sparse.linalg.eigsh.

from pylibraft.solver import eigsh

[lowener]

/ok to test

[lowener]

/ok to test

[cjnolet]

/ok to test

[lowener]

/ok to test

[lowener]

/ok to test

[cjnolet]

Here too- please consolidate this list by using device_coo_matrix_view.

[cjnolet]

We need to get this PR in because code freeze is looming so I'm merging this now. @lowener can we please follow-up by exposing the device_coo_matrix_view in the public API insted of the individual arrays? It should be a 1-2 line change.

[cjnolet]

/merge

[lowener]

@cjnolet I think that the name of the parameters (rows, cols, values) is subject to confusion, and should have been instead (indptr, indices, values). The Lanczos solver only supports CSR format, not COO. There is no conversion from COO to CSR happening in the API point I am adding.
The API that supports the arrays individually and not through the sparse data type was added because the bindings for this sparse data type is not yet present in pylibraft. I created #2504 for this issue