[XeTileToXeGPU] add slm support for the lowering of load_tile/store_tile

[chencha3]

This PR adds lowering support for optimal SLM access. The optimal SLM access is defined as:

tileShape[1] % 16 == 0
slmShape[0] % tileShape[0] == 0
slmShape[1] % tileShape[1] == 0
TileOffset[0] % tileShape[0] == 0
TileOffset[1] % tileShape[1] == 0

regular tile (tile without order attribute)
tileShape[0] x tileShape[1] % 64 == 0 (in bytes) to work
tileShape[0] x tileShape[1] % 256 == 0 (in bytes) for best performance

transposed tile (tile with order attribute)
tileShape[0] % vnni == 0 to work
tileShape[0] % (8 * vnni) == 0 for best performance.

The shape of all uses (init_tile) of the SLM meet the above conditions at the same time.

store/load_tile on regular SLM tile is lowered into 1D xegpu block store_nd/load_nd.
store_tile on transposed SLM tile is lowered into xegpu store_scatter
load_tile on transposed SLM tile is not supported yet, since we don't have a case requiring it.
update_tile_offset on SLM tileType is not also supported, since we don't have a case requiring it.

It also updated the fallback to skip the optimal SLM access and fixed an issue in WgToSg.

Please review these guidelines to help with the review process:

• Have you provided a meaningful PR description?
• Have you added a test, a reproducer, or a reference to an issue with a reproducer?
• Have you tested your changes locally for CPU and GPU devices?
• Have you made sure that new changes do not introduce compiler warnings?
• If this PR is a work in progress, are you filing the PR as a draft?
• Have you organized your commits logically and ensured each can be built by itself?