Use opmath_t in cusparseSpMM (#72559)
Summary:
Fixes https://github.com/pytorch/pytorch/issues/72234
This allows cusparse to compute FP16 inputs using FP32 internal compute type.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/72559
Reviewed By: mruberry
Differential Revision: D34115889
Pulled By: ngimel
fbshipit-source-id: a101a864d301b6cc5e1e5336d8545d466ead32b1
(cherry picked from commit 629ac57628fe773591cce4f2a23f5e6e2d91af9e)
