Rewrite collect_tokens implementations to use a flattened buffer

Instead of trying to collect tokens at each depth, we 'flatten' the
stream as we go allong, pushing open/close delimiters to our buffer
just like regular tokens. One capturing is complete, we reconstruct a
nested `TokenTree::Delimited` structure, producing a normal
`TokenStream`.

The reconstructed `TokenStream` is not created immediately - instead, it is
produced on-demand by a closure (wrapped in a new `LazyTokenStream` type). This
closure stores a clone of the original `TokenCursor`, plus a record of the
number of calls to `next()/next_desugared()`. This is sufficient to reconstruct
the tokenstream seen by the callback without storing any additional state. If
the tokenstream is never used (e.g. when a captured `macro_rules!` argument is
never passed to a proc macro), we never actually create a `TokenStream`.

This implementation has a number of advantages over the previous one:

* It is significantly simpler, with no edge cases around capturing the
  start/end of a delimited group.

* It can be easily extended to allow replacing tokens an an arbitrary
  'depth' by just using `Vec::splice` at the proper position. This is
  important for PR #76130, which requires us to track information about
  attributes along with tokens.

* The lazy approach to `TokenStream` construction allows us to easily
  parse an AST struct, and then decide after the fact whether we need a
  `TokenStream`. This will be useful when we start collecting tokens for
  `Attribute` - we can discard the `LazyTokenStream` if the parsed
  attribute doesn't need tokens (e.g. is a builtin attribute).

The performance impact seems to be neglibile (see
https://github.com/rust-lang/rust/pull/77250#issuecomment-703960604). There is a
small slowdown on a few benchmarks, but it only rises above 1% for incremental
builds, where it represents a larger fraction of the much smaller instruction
count. There a ~1% speedup on a few other incremental benchmarks - my guess is
that the speedups and slowdowns will usually cancel out in practice.

compiler/rustc_parse/src/parser/item.rs

@@ -116,15 +116,16 @@ impl<'a> Parser<'a> {
             Some(item.into_inner())
         });
 
+        let needs_tokens = super::attr::maybe_needs_tokens(&attrs);
+
         let mut unclosed_delims = vec![];
-        let has_attrs = !attrs.is_empty();
         let parse_item = |this: &mut Self| {
             let item = this.parse_item_common_(attrs, mac_allowed, attrs_allowed, req_name);
             unclosed_delims.append(&mut this.unclosed_delims);
             item
         };
 
-        let (mut item, tokens) = if has_attrs {
+        let (mut item, tokens) = if needs_tokens {
             let (item, tokens) = self.collect_tokens(parse_item)?;
             (item, Some(tokens))
         } else {