Iterative CAGRA build temp memory space for graph

[tfeher]

Tracking issue for PR discussion #1810 (comment)

CAGRA's iterative build method needs to allocate a large temp array to store the matrix.

It is expected that we fit into device memory, but the allocation size is still large (n_rows * graph degree).

Is large_workspace_resource the right memory resource to allocate this?

[tfeher]

Additionally, when allocating temporaries, the larges allocation happens after we already spent time building smaller graphs. It would be good to check whether the large allocation fits before we do any work.