initial torch.compile support (inference only)

[hatemhelal]

This PR add some test cases that use torch.compile and a new module mace.tools.compile that contains some helper utilities for MACE compatibilty with torch.compile

Some of the changes needed include:

• remove the torch.jit.script annotations from the scatter-reduce implementations. This is necessary as the compiled script functions are not compatible with the inductor backend.
• disable the torch script compilation baked into the e3nn project by invoking e3nn.set_optimization_defaults(jit_script_fx=False) ahead of creating the model instance. This can be managed with the disable_e3nn_codegen context manager.
• To minimize graph breaks it appears to be necessary to apply a symbolic simplifier to convert the e3nn irreps objects into simpler python types that the inductor compiler knows how to handle. This is accomplished by decorating classes with simplify_if_compile
• This is all put together with the prepare function which manages creating the model without e3nn codegen and applies the symbolic tracing simplification to registered modules.

Note that the remaining graph break in the compiled inference model is due to using autograd to evaluate the forces. This might be possible to fix but I expect it would be easier to do in another PR.

[hatemhelal]

I experimented with the different torch.compile mode options and measured approximate speedups over native eager mode as:

mode	speedup (a10g)
default	1.4
reduce-overhead	1.6
max-autotune	1.6

To get these values I ran:

pytest -s -k test_inference_speedup

Note, still need to investigate the following warning:

skipping cudagraphs due to complex input striding

which is seen for the non-default modes and I suspect is due to how the input batches are created in the test case.