Replace markdown escaping with a combined md/Rd tokenizer

[hadley]

Stacked on #1900.

This replaces escape_rd_for_md()/unescape_rd_for_md() and the double-escaping pass with an explicit grammar for how Rd and markdown mix in roxygen comments. The two languages only overlap at the backslash, so a single byte-wise C++ scan (src/tokenizeMd.cpp) classifies every backslash-initiated construct:

bs_token := "\" NAME braces*   NAME in verbatim list -> tag + all brace args
          | "\" NAME           other tags: args remain markdown
          | "\" PUNCT          two-character escape, e.g. \% \\ \{
          | "\"                lone backslash (never consumes a backtick)

Each token becomes an inert private-use-area placeholder, so commonmark parses text containing no backslashes at all — no escaping, no un-escaping, no placeholder-collision handling. Tokens are restored during the XML→Rd walk with the escaping appropriate to where they landed: as-typed in regular text, Rd-escaped inside \verb{}/\code{}/\preformatted{}, raw inside generated \Sexpr{} bodies. \[ and \] pass through untokenized, because they're commonmark bracket escapes — that single rule is what used to require the de-duplication special case in double_escape_md().

Highlights:

• The first commit adds characterization tests pinning current behaviour for the tricky interactions (Rd tags inside code spans/blocks, links, tables, HTML, multi-line arguments, backslash escapes) before any machinery changed.
• Rd-tags-inside-indented-code-blocks need no block-level parsing at all: the placeholder lands in a verbatim context and is restored as typed, matching the old behaviour by construction.
• escape_verb() loses its hidden ordering dependency on double_escape_md(); all escaping is now local and explicit.
• Re-documenting roxygen2 itself with the new pipeline leaves man/ byte-identical.
• ~2.3x faster on tag-heavy documents (25KB, 400 Rd tags), ~1.3x on typical small texts; profiling shows remaining time is the commonmark parse and XML walk, i.e. the escaping overhead is gone entirely.
• One deliberate behaviour change (NEWS): \% in markdown text now produces \% in the Rd file instead of \\%, which rendered as a backslash followed by a comment that truncated the line.

🤖 Generated with Claude Code

[hadley]

Follow-up commit fixes the two pre-existing warts the characterization tests documented:

• [bar\] (escaped closing bracket) no longer leaks a link reference definition into the Rd output — reference content must not end with a backslash.
• An Rd tag whose brace argument never closes (e.g. \code{a % b}, where the unescaped % comments out the closing brace) now warns with a hint to write \%, instead of being silently reinterpreted as markdown. Output is unchanged.

Both have NEWS bullets; man/ remains byte-identical when self-documenting.

[hadley]

@gaborcsardi very low priority but this is take two at re-doing the escaping/unescaping. I think it's a lot cleaner now.
It doesn't make much difference performance wise, except if you have large sections with tons of Rd tags.