feat(EXLA): Add support for MLIR compilation

[seanmor5]

Requires elixir-nx/xla#45

There's a lot of other changes that need to be made as well. Metal requires MLIR and not XLA Op output, so I need to figure out how to bridge those

[polvalente]

I believe this should be linked in the respective release for elixir-nx/xla directly, and then you can have something like this coming from there:

nx/torchx/c_src/torchx.cpp

Lines 347 to 355 in 219c23e

	NIF(mps_is_available)
	{
	#ifdef MAC_ARM64
	bool has_mps = at::hasMPS();
	#else
	bool has_mps = false;
	#endif
	return nx::nif::make(env, has_mps);
	}

[polvalente]

For Torchx, this flag is defined here:

nx/torchx/CMakeLists.txt

Lines 46 to 48 in 219c23e

	if(ARM64_SUPPORTED)
	set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DMAC_ARM64")
	endif()

I assume you'd do something similar for :xla

[josevalim]

If we want to keep doing it on Elixir land: we can load the plugin on EXLA.Application. I also agree that ideally we would bundle the plugin with the precompiled XLA binary for macOS.

[seanmor5]

I don't think that's going to work here because the only way to register the plugin as far as I can tell is dynamically through PjRt

[josevalim]

What I mean is that we ship the plugin with XLA and in here we point to XLA priv dir to load it. Would that work?