[TRITON-BP]: Lack of support for ExpandDimOp

[etiotto]

Failing test:

/home/jovyan/intel-xpu-backend-for-triton/test/Triton/Intel/RaiseToBlockPointers/addptr_for_expand_ptr.mlir:21:13: remark: TritonRaiseToBlockPointer: ExpandDims Ops in loops are currently not supported
    %_ptr = scf.for %i = %c0 to %c12 step %c3 iter_args(%ptr = %2) -> (tensor<256x!tt.ptr<bf16>>) {

[etiotto]

This happens when tt.expand_dims is in the body of an scf.for loop and references a block ptr passed into the loop as a init_arg.
For example:

    %6 = tt.make_tensor_ptr %arg0, [%c0_i64], [%c1_i64], [%c1024_i32] {order = array<i32>} : <tensor<256xbf16>>
    %7 = scf.for %arg1 = %c0 to %c12 step %c3 iter_args(%arg2 = %6) -> (!tt.ptr<tensor<256xbf16>>) {
      %8 = tt.make_range {end = 256 : i32, start = 0 : i32} : tensor<256xi32>
      %9 = tt.expand_dims %8 {axis = 1 : i32} : tensor<256xi32> -> tensor<256x1xi32>
      %10 = tt.broadcast %9 : tensor<256x1xi32> -> tensor<256x256xi32>
      %11 = tt.make_range {end = 512 : i32, start = 256 : i32} : tensor<256xi32>
      %12 = tt.expand_dims %11 {axis = 0 : i32} : tensor<256xi32> -> tensor<1x256xi32>
      %13 = tt.broadcast %12 : tensor<1x256xi32> -> tensor<256x256xi32>
      %14 = arith.addi %10, %13 : tensor<256x256xi32>

     %x = tt.expand_dims %ptr {axis = 1 : i32} : tensor<256x!tt.ptr<bf16>> -> tensor<256x1x!tt.ptr<bf16>>

      %15 = tt.splat %c3_i32 : i32 -> tensor<256xi32>
      %16 = tt.advance %arg2, [%c3_i32] : <tensor<256xbf16>>
      scf.yield %16 : !tt.ptr<tensor<256xbf16>>
    }

Here %arg2 is a block ptr. The expand_dims operation yielding %x references it. We do not know how to rewrite that operation given that its argument %ptr is incremented at each loop iteration.

I think this pattern doesn't happen too often. So for now we can keep this a limitation. That is, when the scf.for loop contains an operation that references a rewritten ptr, but that operation cannot accepts a block ptr as its operand. We give up transforming the loop.