Add temp storage alignment awareness

[tpn]

Numba CUDA recently introduced the ability to specify an alignment=N keyword arg to the cuda.(shared|local|.array() helpers in this PR. We need to update our cuda.cooperative Algorithm implementation to instrument returned callables with an additional temp_storage_alignment attribute in addition to the existing temp_storage_bytes attribute.

Specifically, temp_storage_alignment is equivalent to alignof(Alg::TempStorage), as temp_storage_bytes is equivalent to sizeof(Alg::TempStorage).

An example of a concrete issue this will fix: create two temp storage arrays in a kernel; the second one gets a byte alignment, so if you're trying to use a cuda.coop primitive with, say, a float, which needs 4-byte alignment, the kernel will trap.