[TritonStructuredOps] Separate MakeTensorOp into three separate operations for block pointers, structured pointers, and split pointers.

[red1bluelost]

The MakeTensorOp can become an interface like LinalgOp while each style of pointer has a dedicated operation. This will avoid operands being ignored for one style but not another. This helps make the code clearer to a developer while still allow the shared behaviors through the interface.

[Nullkooland]

Hi, if you are refactoring this, could you use per-dim offsets instead of scaled offsets in tts.make_tptr? See #208 and #203.

[enjustli]

Hi, if you are refactoring this, could you use per-dim offsets instead of scaled offsets in tts.make_tptr? See #208 and #203.

i thought we can reserve tts.make_tptr && tts.load/store, add hardware-specific pass to resolve this problem. i.e., disable structured-to-memref pass