Support tiling and swizzle transforms in GPU Interpret Mode.
This change saves the ref transforms as a param in `pallas_call_p` and `run_scoped_p`, so that GPU Interpret Mode can use them to determine the untiled shape for kernel args and run_scoped allocations. GPU Interpret Mode buffers are then allocated with untiled shapes and we ignore untiling and unswizzle transforms when interpreting the body of a kernel or a run_scoped.
PiperOrigin-RevId: 928764383
