mbe: compile rule RHS into a pre-flattened transcription template [clap_derive -0.5%, tt-muncher -1.8%]

[xmakro]

Follow-up to #57, stacked on #82. Transcription walked the RHS tree token by token on every expansion of a rule: one dispatch, one span mark, and one buffer push per literal token, plus a metavar normalization per occurrence and a full tree re-walk for every sequence lockstep check.

This lowers each rule RHS once (lazily, on first expansion) into a compiled template:

• Runs of literal tokens, including whole literal delimited groups, are pre-encoded as flat entries with def-site spans. Transcription copies each run into the output buffer, rebases depths and match indices, and marks the copied spans in place. The per-token interpreter dispatch is gone for these.
• Metavar occurrences carry their MacroRulesNormalizedIdent precomputed.
• The sequence lockstep size check walks a per-rule precollected list of the metavars in the sequence subtree instead of re-walking the tree on every sequence entry.
• Definitions that are never invoked pay nothing (the template is built behind a OnceLock).

Diagnostics are preserved: the sequence segments keep the spans and metavar ident lists the lockstep and no-repeatable-vars errors need, and the close-bracket spacing hack (rust-lang#114571) is replicated in the pre-encoded runs.

Performance (ThinLTO stage2, jemalloc, instructions:u, full scenario, measured on top of the perf stack; A/B over 10 crates x Check/Debug, base includes #82):

• clap_derive check -0.52%, html5ever check -0.46% / debug -0.38%, syn check -0.27%, serde -0.16%
• tt-muncher check -1.81%, debug -1.17% (about -4% cumulative with mbe: cache metavar resolution across repetition iterations [tt-muncher -2.2%] #82)
• geomean -0.30%, 6 cells improved >= 0.25%, 0 regressed

Correctness: stage2 build green; x check compiler green on this branch standalone; full tests/ui run has a failure set byte-identical to the pre-existing baseline (21301 passed, zero new diffs).