Investigate the new tensor descriptor API

[mfrancepillois]

OpenAI has improved the way structured memory access is handled.

The PR : https://github.com/triton-lang/triton/pull/4916 cleans-up, extends the triton dialect with new operations and improves the way TMA descriptors are handled by triton.

As significant changes result from this PR, we should investigate if and how these memory accesses using tensor descriptors could be transformed into block pointers.

[mfrancepillois]

The information needed to create a block pointer are:

• Base: pointer to parent tensor​
• Shapes: parent tensor shape​
• Strides: parent tensor strides​
• Block_shape​
• Offsets: block offsets​
• Order: block order

The new proposal for Tensor descriptor allows users to create a tensor descriptor on the device, which will be lowered into a TMA descriptor. This PR extends the Triton Dialect with a new operation MakeTensorDescOp. This operation contains the following information:

• Base: pointer to parent tensor​
• Shapes: parent tensor shape​
• Strides: parent tensor strides​
• Block_shape​ Combining these pieces of information with the offsets included in the experimental_descriptor_{load/store} operations, could allow us to have all the required information to translate Tensor descriptor into block pointer. ​ A possible solution could be to add a pass to the front of the Intel pipeline that:​
  1. Search for load/store Ops using a ptr created by a MakeTensorDescOp (at some point should be a specific type, see https://github.com/triton-lang/triton/blob/837308f780066a6606df4362a9648f9ce55c625b/include/triton/Dialect/Triton/IR/TritonOps.td#L961)​
  2. Create a MakeTensorPtrOp just before the load/store with information collected from the MakeTensorDescOp + offsets from the load/store operation​ Notice that, in that case, the MakeTensorPtrOp cannot be moved (or must be moved with care => cannot be moved before the offset computation)​

Possible alternative: keep the MakeTensorDescOp as it is in the pipeline and implement only our XPU-specific lowering that should use 2D block operations.