Fixing several bugs in Halide's AMX support

[yihozhang]

This PR exposes and fixes a number of issues with Halide's current AMX support:

• Ensures that the modulo in RHS matrix is correct (tiled_matmul_wrong_modulo.cpp)
• Removes the stride in the 1D case, since there will be just one row and the stride does not really matter.
• Checks that the RHS tile to be loaded is contiguous in its rows (tiled_matmul_wrong_layout.cpp)
• Ensures that the tiling of RHS is 4 / element_width (tiled_matmul_wrong_tiling.cpp, tiled_matmul_wrong_pattern.cpp)
• Swaps tile_x and tile_y in the 3D case
  • this was not found in the first place since tile_x=tile_y always holds in the original tests

Not yet addressed

• The second child to the Add node in the matcher is ignored.
• support cases where tile_x = 1 ortile_y=1
• Load's predicate is not checked
• Make clear that the optional Broadcast pattern should only occur for the 1D and 2D case