Test for compile of compile (and compile of existing compiled thunk)

[wsmoses]

Down the line I imagine making sure this works properly to be critical.

For example suppose one uses Reactant inside of scimlsensitivity to compile the function. For a generic code this is fine. However, suppose someone has a lux NN inside of their function, using Reactant (cc @avik-pal). Now they may have a compile of a compile as a result.

[mofeing]

This is a common problem our users encounter. Note that we already cache the Expr of a Thunk, but we don't yet cache the MLIR after the tracing, which we probably need to make this work.

[wsmoses]

yeah Enzyme itself does a similar trick where it saves the IR in a map of "generated fn pointer -> corresponding IR", and then it replaces a call to the fn pointer with injecting that IR.

We'd presumably need something somewhat similar but for thunks and mlir