Opportunistically split != to successfully parse never type

[fmease]

Accept code like let _: != loop {}; or let _: fn() -> != || loop {}; (previously: syntax error). Similar to sth. like #51068.

There's no inherent need to update can_begin_type from what I can tell, since macro matcher $ty:ty = already accepts != (edit: this is not true actually, my bad, I don't know what snippet I tested to come to that conclusion) and since no other current direct/transitive callers would 'benefit' from doing that either, so I didn't do it. It's a conservative choice to avoid odd consequences (concrete example: it would widen the follow set of :vis to include !=).

[fmease]

r? petrochenkov or reassign

[fmease]

Maybe add this, too:

#![feature(default_field_values)]

struct X<T> {
    w: fn(T) -> ! = |_| loop {},
    s: fn(T) -> != |_| loop {},
}

Doesn't really matter tho.

[petrochenkov]

Yeah, let's add this one too.

[petrochenkov]

It will probably work if the support is added to break_two_token_op.

[traviscross]

As this changes the language grammar, this will need a lang FCP. If you could talk a bit about the analysis and motivation here for the nomination, that would be helpful.

This change would allow, e.g.,

let _: fn() -> != || loop {};

to be accepted both syntactically and semantically.

It'd be good for us to find a way to document this in the Reference as well.

cc @rust-lang/lang-docs

(H/t to our lang advisor @ehuss for raising this.)