Support canUpcastAcc in TileAndFuse pipeline when using intrinsics

[nirvedhmeshram]

The TileandFuse Matmul config heursictics has this code snippet

  if (!schedule) {
    // Then try again by allowing upcasting accumulator.
    schedule = deduceMMASchedule(
        problem, intrinsics, seeds, maxSharedMemoryBytes, targetSubgroupSize,
        transposedLhs, transposedRhs, /*canUpcastAcc=*/true,
        /*mustBeAligned*/ mustBeAligned, doCPromotion);
  }

However, it cannot actually upcast becuase we will fail to make a multi-mma op here
https://github.com/iree-org/iree/blob/main/compiler/src/iree/compiler/Codegen/Dialect/GPU/Transforms/Transforms.cpp#L446-L447
as we dont have upcast + multi_mma + down cast ability.

VectorDistribute does this after vectorization here with the LLVMGPUCastTypeToFitMMA pass.

For TileAndFuse I think ideally we would want to do something similar when we make the multi_mma op which I did in this branch
However, doing it early like this introduced two issues that I noticed at the FuseAndHoist pass

1. The extf leads to a materialization of empty+fill+extf which should be just an empty
2. The truncf doesnt seem to be tiled along with the multi_mma
   Here is a dump

The end result is that it failed to bufferize.