Moved Pallas GPU lowering registartion code into a separate submodule

This makes the layout similar to the one we use in Pallas TPU.

PiperOrigin-RevId: 610411536

jax_triton/triton_lib.py

@@ -159,7 +159,7 @@ def compile_ttir_to_ptx_inplace(
     cuda_options: cb.CUDAOptions,
     device: int = 0,
     device_type: str = "cuda",
-) -> Tuple[str, str, int, int]:
+) -> Tuple[str, str, int, int, Any]:
   compute_capability = triton_kernel_call_lib.get_compute_capability(device)
   if cuda_options.debug:
     print(ttir)