Properly stall coroutine witnesses in new solver

[compiler-errors]

Fixes rust-lang/trait-system-refactor-initiative#82.

Using an infer var for the witness meant that if we constrain the infer var during writeback and then try to normalize during writeback, after the coroutine witness has been plugged into the coroutine type - which we do with the new solver - we may encounter a query cycle due to trying to fetch the coroutine witness types.

This PR changes the AnalysisInBody typing mode to track all coroutines being defined by the current body during typeck, and forces any auto trait and Copy obligations that would require fetching the hidden types of these coroutines to be forced ambiguous. This also introduces a new proof tree visitor which detects which obligations should be stalled due to bottoming out in one of these ambiguous obligations, so we can re-check them after borrowck (as is done with the old solver).

This PR shouldn't have functional changes, but post-mortem seems to have introduced a perf regression. Looking at the code, I don't see much of a reason why this would be the case. We don't call the new query when the old solver is active, nor should be be visiting any of this new unstalling code in the old solver.

r? lcnr