soundly supporting specialization

[lcnr]

Supporting specialization in the new solver is challenging.

We only add alias bounds for projections which are well-formed due to a ParamEnv candidate (and add a special-case for builtin traits which act as if they are specialized). This is necessary for soundness (#6). It causes us to fail to prove <T as Trait>::DefaultAssoc: Sized with the new solver. Because specialization is used by std we may also allow alias bound candidates for specializing impls, even if unsound.

Other issues to be added here.

[lcnr]

this is outdated, think about this again