Do not eagerly reject inference vars when trying to resolve method calls.

[oli-obk]

Reintroduces the changes reverted in #126258 and fixes the regression in the old solver.

The new solver will still fail on

pub trait Parser<E> {
    fn parse(&self) -> E;
}

impl<E, T: Fn() -> E> Parser<E> for T {
    fn parse(&self) -> E {
        self()
    }
}

pub fn recursive_fn<E>() -> impl Parser<E> {
    move || recursive_fn().parse()
    //[next]~^ ERROR: type annotations needed
}

because while type checking the method call .parse(), when looking for candidates, we end up not finding any, as the only information we have is that there's an unresolved AliasTy(opaque, infer_var) floating around while looking for a method call on infer_var.

This obviously seems fixable by looking at said obligation, but I'd prefer to tackle that separately (I got half an impl locally, but there are multiple paths forward that we should discuss)

r? @lcnr @compiler-errors

I'm going to split out the effect-free cleanups and the inference changes to separate PRs, so we can review them in isolation, but with the goal of enabling this PR to land

[lcnr]

want to take some time to look at when exactly we don't error if there's an auto-deref step resulting in an infer var, will get to this next week hopefully

[oli-obk]

errors like this are pretty useless and should be fixed before landing the PR

[oli-obk]

This PR fixes the time inference regression caused by #99969

I can see how that happens, but it's best effort, as with most inference errors or fixes we can't make long term guarantees...

[oli-obk]

as per the comment above, this diagnostic needs to be fixed again