Speed up traced_type by conversion to generated function

[wsmoses]

Can be done analagously to Enzyme.active_reg_inner. This means we will have type inference return a constant type here (and make traced_type and the rest much faster/infer). Moreover we can leverage a global cache for all type queries. It will be invalidated upon a new def of traced_type_inner.

This should speed up all calls to traced_type.

It does require traced_type to be computable as a function of type arguments.

@avik-pal do we need the sharded value for traced_type_inner or does the sharded type suffice?

[avik-pal]

Currently only needed for the PJRT number of devices which depends on the exact value, though we could evaluate that before calling the traced_type and pass in a Val

[wsmoses]

Yeah if we can move that to the type domain, I can handle the generated conversion