Refactor inlining to allow re-use in more sophisticated inlining passes

[Keno]

The inlining transform basically has three parts:

1. Analysis (What needs to be inlined and are we allowed to do that?)
2. Policy (Should we inline this?)
3. Mechanism (Stuff the bits from one function into the other)

At the moment, we already separate this out into two passes:
Analysis/Policy (assemble_inline_todo!) and Mechanism (batch_inline!).
For our needs in base, the policy bits are quite simple (how large
is the optimized version of this function), but that policy is
insufficient for some more sophisticated inlining needs I have
in an external compiler pass (where I want to interleave inlining
with different transforms as well as potentially run inlining multiple
times). To facilitate such use cases, this commit optionally splits
out the policy part, but lets the analysis and mechanism parts be
re-used by a more sophisticated inlining pass. It also refactors
the optimization state to more clearly delineate the different
independent parts (edge tracking, inference catches, method table),
as well as making the different parts optional (where not required).
We were already essentially supporting optimization without edge
tracking (for testing purposes), so this is just a bit more
explicit about it (which is useful for me, since the different
inlining passes in my pipeline may need different settings).

For base itself, nothing should functionally change, though
hopefully things are factored a bit cleaner.