[LinalgToXeGPU] Support squeezable any-D memrefs

[dchigarev]

This PR allows any-D memrefs to be used in linalg-to-xegpu pass as long they can be squeezed to a 2D memref through leading dimensions. Example:

// allowed
!chunk_type = memref<1x1x8x16xf16, strided<[512, 128, 16, 1], offset: ?>>
linalg.mul ins(%0, %1 : !chunk_type, !chunk_type) outs(%2 : !chunk_type)

// NOT allowed
!chunk_type = memref<1x8x1x16xf16, strided<[512, 128, 16, 1], offset: ?>>
linalg.mul ins(%0, %1 : !chunk_type, !chunk_type) outs(%2 : !slm_chunk)

The linalg-to-xegpu logic considers a linalg operation that can be lowered to xegpu and checks the operand shapes. The old logic refused all linalgOps that have non-2D operands. The new logic additionally checks if a non-2D operand can be squeezed to a 2D and allows them if they can. It then squeezes the non-2D operand using memref.subview and replaces old operands with new 2D memrefs. Example:

// input:
!chunk_type = memref<1x1x8x16xf16, strided<[512, 128, 16, 1], offset: ?>>
linalg.mul ins(%0, %1 : !chunk_type, !chunk_type) outs(%2 : !chunk_type)

// output:
!chunk_type = memref<1x1x8x16xf16, strided<[512, 128, 16, 1], offset: ?>>
!squeezed_type = memref<8x16xf16, strided<[16, 1], offset: ?>>

%0_sq = memref.subview %0[0, 0, 0, 0] [1, 1, 8, 16] [1, 1, 1, 1] : !chunk_type to !squeezed_type
%1_sq = memref.subview %1[0, 0, 0, 0] [1, 1, 8, 16] [1, 1, 1, 1] : !chunk_type to !squeezed_type
%2_sq = memref.subview %2[0, 0, 0, 0] [1, 1, 8, 16] [1, 1, 1, 1] : !chunk_type to !squeezed_type
linalg.mul ins(%0_sq, %1_sq : !squeezed_type , !squeezed_type ) outs(%2_sq : !squeezed_type )

[dchigarev]

the logic was taken from XeGPUFoldAliasOps upstream pass

[kurapov-peter]

Some minor comments, fine otherwise