Support TMA with 64-bit indexing

[jacobhinkle]

This PR removes the check that IndexType is Int32 when using TMA.

When TMA is enabled we have a coordinate array specifying the data pointer of the global memory array and the box coordinates. The global array might be larger enough to require an int64_t to linearly index its last element, in which case we would use int64_t indexing. However, since it is possibly a multidimensional array, it is still possible that the multidimensional coords of all TMA boxes can still be expressed with int32_t. These are the cases we want to cover for matmul since the input is 2D and we don't commonly expect each individual dimension to be larger than 2^31.

Currently I am just static casting the box dims to int32_t in memory.cu. We should additionally verify that none of these coords will be larger than the capacity of int32_t. We could handle that more broadly actually by writing a more sophisticated analysis that takes every indexing expression (including TMA box coords) in the Fusion and proves bounds for it. This is trickier than what we currently do which is just look at all TensorViews and determine the position of the furthest strided element within that tensor, but it might reveal that we use int64_t in many cases where it's unnecessary (such as matmuls).

Note that it's still possible to have a box coordinate that cannot be expressed with a 32-bit index, in which case we should not allow TMA.

TODO: we should check the above condition in the core heuristic where we have access to problem sizes.

Fixes #3595

[jacobhinkle]

!test

[rdspring1]

LGTM.