Build callgraph of guppy functions, use to calculate effects of unannotated funcs

[acl-cqc]

This fits inbetween checking (+monomorphization) and compilation, but we'll need to modify checking (specifically check_call and synthesize_call) to build the callgraph.

The intervening analysis phase uses a graph library to identify cycles (connected components); all functions in a CC will have the same effects. Effects of unannotated functions can then be calculated by a tree traversal (union of effects in the body of a function/cycle plus those in any callee outside of the cycle).

Whether we should error on finding a call from a max_effects= function to one that's unannotated is debateable - we could add a strict_effects=<bool> flag but we can probably allow this and instead enforce that every @guppy.declare is explicitly annotated along with every library export (otherwise it's very easy to introduce breaking changes).

[acl-cqc]

Looking at #1723, class EffectLimitDecl currently stores the effect of an explicit effects= annotation; Context stores an EffectLimitDecl | None (note in how few places Context is constructed- tracing and check_global_func_def - we'll also want to start making a new Context in check_nested_func_def.)

So, replace the EffectLimitDecl in Context with "source node in callgraph", and where _check_effects checks legality of calls, instead just record all the calls. (We could raise errors for calls where both caller and callee are annotated but this is probably just going to repeat code.) That is, this builds the callgraph.

Callgraph nodes can then have an associated EffectLimitDecl to record any explicit annotation. After the callgraph is built and then analysed (i.e. SCCs), we'll then need to check calls from annotated functions: certainly to unannotated functions, i.e. if the analysis has figured out the callee has too many effects; probably to all functions (i.e. also annotated ones), as that's more uniform than dealing with these first/earlier.