[Pallas/Mosaic GPU] Expose `plgpu.wgmma_accumulator_load`.
This function allows loading from a WGMMA accumulator while controlling the
amount of `wgmma` instructions that are allowed to remain in flight before
dereferencing occurs.
If the `wait_n` parameter is set to `None`, no synchronization is done.
PiperOrigin-RevId: 858186473
