The existing `InferCtxt` may already have in-progress projections
for some of the predicates we may (recursively evaluate). This can
cause us to incorrect produce (and cache!) `EvaluatedToRecur`, leading
us to incorrectly mark a predicate as unimplemented.

We now create a fresh `InferCtxt` for each sub-obligation that we
'speculatively' evaluate. This causes us to miss out on some
legitimate caching opportunities, but ensures that our speculative
evaluation cannot pollute any of the caches from the 'real' `InferCtxt`.
The evaluation can still update *global* caches, but our global caches
don't have any notion of 'in progress', so this is fine.

Fixes rust-lang#90662