Simplify `#[derive(BuiltinTrait)]` expansion wrt. lifetimes

[Centril]

For example, given:

#[derive(Clone)]
struct A<'a>(&'a u8);

we generate:

impl<'a> Clone for A<'a> { ... }

but we should be able to generate:

impl Clone for A<'_> { ... }

This might lead to perf improvements by generating less tokens.

cc @zackmdavis @scottmcm @eddyb @petrochenkov
cc #61824

[eddyb]

If we need to generate any bounds, at all, this will be tricky as we'd have to check which lifetime names would show up in the bounds. Other than that, it seems straight-forward.
One thing we should make sure first is that we don't use A<'a> anywhere today other than the impl header (and that we're using Self everywhere else).

[jyn514]

Is this useful? It seems awfully complicated, and I don't think the lexer is the slow part of the compiler ...

[petrochenkov]

This kind of leads to a philosophical question - what is the purpose of an issue tracker? Should it track potentially nice ideas that nobody is going to implement, not even the author?

I'd personally close this, if someone is ready to invest work into proving usefulness of this idea and measure performance, then a PR is welcome, but otherwise we should not track this as an open issue.