asahi: Turn agxdecode into a shared library usable by the hypervisor

Because why not~ ✨

The main refactoring/difference here is the GPU memory reads no longer grab direct pointers into allocs, but always copy data into pre-allocated buffers. This is, of course, necessary with the hypervisor since it doesn't have a mapped view into GPU memory.

Other than that it's just adding a handy interface that Python can use with ctypes.

Based on the G14X MR because that also touched decode.