[Mosaic GPU] Add async_copy_sparse_metadata_smem_to_tmem dialect lowering.
No test in dialect for this one, because we do not support sparse tmem layouts for cs.IsTransferable. This implies that in a single test we cannot go TMEM -> Regs -> GMEM to verify numerics. In a subsequent CL, pallas lowering is tested via unblocking sparse_mma test.
PiperOrigin-RevId: 884429087