[Backend] Refactor sharedToDotOperandMFMA lowering

[zhanglx13]

This PR

• Implements https://github.com/ROCmSoftwarePlatform/frameworks-internal/issues/5237
• Supports mfma16 and mfma4 in the fast path
• sets maxPhase=4 for mfma4. This is because the nonKDim of mfma4 is less than 16.
• Choose warpsPerTile according to mfmaLayout's nonKDim
• cleans up format

Before we enable vector access to LDS for non k-major tensors, here are our conventions about LDS accesses:

• LDS write
  • For k-major tensors, swizzling is enabled, i.e. a vector of elements are swizzled in LDS. ds_write are vectorized
  • For non k-major tensors, a.k.a. mn-major tensors, swizzling is disabled, i.e. elements are written to LDS in their original order. The LLVM backend usually vectorizes ds_write for us, though we write one element at a time.
• LDS read
  • For k-major tensors, normal path is taken since swizzling is enabled.
  • For non k-major tensors, fast path is taken, since the mapping from an element's coordinates in the tensor to its offsets in the LDS is linear.

Note

• If the tensor is k-major, we assume swizzling is enabled. Therefore, we don't allow/test the case where swizzling is disabled for k-major tensors.

[zhanglx13]

but I suggest to fallback to normal path if operand is non-k major.

@alefimov-amd I assume you meant fallback to normal path even if swizzling is disabled for k-major tensor.