Improve token stream pretty printing.

[nnethercote]

Specifically:

• No space before or after ::, e.g. a::b instead of a :: b.
• No space between ! and (, e.g. foo!() instead of foo! ().
• No space between ! and [, e.g. #![...] instead of #! [...].
• No space before :, e.g. a: u32 instead of a : u32.
• No space before ;, e.g. struct A; instead of struct A ;.

r? @petrochenkov

[rust-highfive]

r? @cjgillot

(rust-highfive has picked a reviewer for you, use r? to override)

[nnethercote]

The a::b change is a necessary part of #96724, without it some doctests fail. While fixing that I saw some other easy improvements. I figure it's worth pulling these out on their own because #96724 is going to take a while.

[nnethercote]

r? @petrochenkov

[fbstj]

what is the FIXME for?

[petrochenkov]

Originally spaces were always printed because proc macros worked through pretty-printing and reparsing, and spaces prevented tokens from being glued together and forming other tokens, like : : -> ::.
Note that in proc macros spaces are significant for all punctuation tokens, not only those that can form multi-character operators in Rust syntax, e.g. !+ and ! + can be treated differently, similarly to < < and << in Rust syntax.

Then we started skipping some spaces to avoid breaking code (#63897).

Eventually people just started skipping them to beautify the pretty-printed code, which is incorrect in theory.
But it may be less important now when proc macros very rarely have to go through pretty-printing and reparsing.

[petrochenkov]

Ideally we would preserve jointness information for all tokens and then use it during pretty-printing (#76285 went into that direction, but was reverted later due to some issues and never re-landed).

[petrochenkov]

To keep documentation links working with #96724 it should be enough to elide spaces in cases :: IDENT and IDENT ::, so I'd personally prefer to limit this PR to that.

Although ! ( and ! [ should be fine too because they can never be joined.