[llvm] Reenable sched mfma rewrite (PR #180751)

[Jeffrey Byrnes via llvm-commits]

jrbyrnes wrote:

In https://github.com/llvm/llvm-project/issues/169552 we are seeing some benefit to enabling this rewrite (remove a bunch of accvgpr_read/write + v_mov in loop), but we introduce some redundant copies in the epilogue.

See - https://godbolt.org/z/Y8G1bd98f

I'm seeing stuff like this in the epilogue

```
        v_accvgpr_read_b32 v0, a212
        v_accvgpr_read_b32 v4, a212
        v_accvgpr_read_b32 v2, a214
        v_accvgpr_read_b32 v3, a215
        v_accvgpr_read_b32 v5, a213
        v_accvgpr_read_b32 v6, a214
        v_accvgpr_read_b32 v7, a215
        v_accvgpr_read_b32 v1, a213
        v_cvt_pk_bf16_f32 v7, v6, v3
        v_accvgpr_read_b32 v2, a212
        v_accvgpr_read_b32 v3, a213
        v_accvgpr_read_b32 v4, a214
        v_accvgpr_read_b32 v5, a215
```

I think this is a known issue, but I'm wondering if this is actively being worked on / if this is planned to go in along with the reenable?

https://github.com/llvm/llvm-project/pull/180751