[Unity][CUTLASS] Support out_dtype = "float32" for FasterTransformer kernel

[masahi]

The FT kernel requires the output pointer type to be fp16, so we haven't supported offloading matmul with out_dtype="float32". But if such matmul is immediately followed by a cast to fp16, we can match and replace the whole matmul -> cast with the FT kernel.

Such case arises if a user wants to do fp32 accum during matmul. It turns out the FT kernel always does fp32 accumulation internally regardless of the output dtype, so we are not losing any precision by offloading matmul(..., out_dtype = "float32") -> cast to the fp16 A - int B FT kernel.

@vinx13 @sunggg @yzh119

[tvm-bot]

Thanks for contributing to TVM! Please refer to the contributing guidelines https://tvm.apache.org/docs/contribute/ for useful information and tips. Please request code reviews from Reviewers by @-ing them in a comment.

• cc @quic-sanirudh _{See #10317 for details}

_{Generated by tvm-bot}