Currently it just unconditionally allocates it in the arena.

For a "Clean Check" build of the the `packed-simd` benchmark, this
change reduces both the `max-rss` and `faults` counts by 59%; it
slightly (~3%) increases the instruction counts but the `wall-time` is
unchanged.

For the same builds of a few other benchmarks, `max-rss` and `faults`
drop by 1--5%, but instruction counts and `wall-time` changes are in the
noise.

Fixes rust-lang#57432, fixes rust-lang#57829.

This commit fixes a bug introduced by rust-lang#55937 which started checking user
type annotations for associated type patterns. Where lowering a
associated constant expression would previously return a
`PatternKind::Constant`, it now returns a `PatternKind::AscribeUserType`
with a `PatternKind::Constant` inside, this commit unwraps that to
access the constant pattern inside and behaves as before.

In particular, the table entries (associated with type-variables
created during the probe) must persist as long as the candidates
assembled during the probe. If you make a nested probe without
creating a nested `ProbeContext`, the table entries are popped at the
end of the nested probe, while the type-variables would leak out via
the assembled candidates attached to `self` (the outer
`ProbeContext`). This causes an ICE (*if you are lucky*)!

These aren't fixed by this PR, but were broken in a few older attempts
at it. Make sure they don't regress.

…bini