fix: Fix matching of doc comments in MBEs

[ChayimFriedman2]

They should be ignored, but only if they stay in the original representation as tt::DocComment. Which first means we need to author tt::DocComment, then handle the cases where it's not preserved: in MBE input, and when proc macros see it (but lazily; not when passed to proc macros, only when they inspect it - i.e. only when it's converted to the proc-macro-bridge's lossy representation.

Invisible delimiters of MBE, one of the most frequent origins for bugs we have, also work in the same way (both originate from how rustc does not separate between its AST and macro input), so it's a precedent more important than just this edge case.

I don't think we have a way to test the interactions with proc macros (this requires a real proc macro server with the bridge), but I tested all gory details manually with a built proc macro server.

This requires a tiny adaptation from the RustRover folks (we should probably notify them): the serialization of FlatTree has changed slightly (even in the JSON format) for version >= DOC_COMMENT_LEAF. They can send (and receive) an empty doc_comments list if they do not send non-desugared doc comment token trees (which they do not, currently).

Specifically, in those versions the schema went from:

interface FlatTree {
    subtree: u32[],
    literal: u32[],
    punct: u32[],
    ident: u32[],
    token_tree: u32[],
    text: string[],
}

Into:

interface FlatTree {
    pre_doc_comment_leaf: {
        subtree: u32[],
        literal: u32[],
        punct: u32[],
        ident: u32[],
        token_tree: u32[],
        text: string[],
    },
    doc_comments: u32[], // Can be empty for RustRover.
}

Fixes #23088.

This sits on the top of #23152 which is required for this.