Refactor race warnings (accesses) to use global invariant

[sim642]

This is a continuation of #368 for race warnings, which aren't done the usual way: they don't appear during individual transfer functions.
These changes could apply to master directly, but since they're just necessary for interactive race warnings, this PR is against interactive, where there are incremental warnings at all.

Changes

1. Extract access handling from mutex analysis to separate access analysis.
2. Collect accesses into global invariant using sideg and deeply nested map domains instead of using a OCaml global deeply nested hashtable. Having this in the actual constraint system and its solution makes it easier to do incremental accesses/race warnings.
3. Allow sideg during postsolving to modify the solution. This is a massive hack, but since accesses are only collected at the end, this is the only way to have them added into the solution. To make verify happy, leq of the access analysis global domain is constant true (it's only collected during postsolving anyway). The alternative would be to expose a pair of domains through all the specs, constraint systems and solvers to allow the second warning-collecting component to be handled separately.
4. Add WarnGlobal query to emit warnings for globals at the end of solving. Since race warnings are based on the accesses collected into the global invariant during postsolving, this needs to be done per-global after postsolving.

TODO

• This more-or-less works with incremental, but type-based accesses go under a special __NONE__ varinfo, which is initialized to G.bot (), and currently set_start overwrites with this during incremental load. Hopefully Incremental TD3: fix start variables #393 fixes this.

[sim642]

Merging into the branch interactive for better experimentation.