[AMDGPU] Insert inliner anchor earlier
Add a new hook for inserting passes right after the last DummyCGSCC pass
and use it to insert the anchor. This changes the last FunctionPass
manager to be an inlining pass manager, thus preserving some of the
analyses that might be computed before the inliner and used after it (to
be fair that's never going to be a lot of analyses, since inlining is
pretty plastic, but at least some of the IR-level analyses that have
absolutely no reason to change can be computed only once).
This is how I originally designed the code, but I don't feel like I have
a good name/abstraction for this exact point in the pipeline, hence the
separate patch.
