Infer extension-deltas for FuncDefn's

[acl-cqc]

We need to decide whether this is really worth it, i.e. is it too painful (and not enough of a good thing anyway) to require FuncDefn authors to declare what extensions they use. But there is at least some utility in that it's optional anyway.

This could be done with the same interface to that for DFGs/CFGs/etc. (using TO_BE_INFERRED: ExtensionId as per #1195), but the algorithm needs to be significantly more complicated:

• Build a (directed) call graph of functions, each node has a lower-bound ExtensionSet of the non-Call leaf ops
• Identify cycles - all functions on a cycle must have the same ExtensionSet, so "merge metas" (replace cycle with single node with lower-bound being the union)
• Solve the unknowns - each is just union over the successors (things it must contain)

This is very roughly what was in old #1137