mbe: match eligible arms directly on the flat token entries [tt-muncher -6%, clap_derive -1.5%]

[xmakro]

Follow-up to #57, stacked on #84. The matcher no longer drives a full Parser over the input for arms it can handle directly.

When no tracker observes matching and an arm's captures are all resolvable straight off the flat token entries (tt, ident, lifetime), matching steps a lightweight flat input instead of a parser:

• advancing is an entry load instead of a Parser::bump (no prev_token/spacing bookkeeping)
• tt fragments become buffer slices directly, with no black-box parser round trip through parse_nonterminal
• ident and lifetime captures read the token in place; nonterminal_may_begin_with already vets the token with the same predicates, so these are infallible
• Eof gets the same last-token span tweak Parser::bump applies, and invisible-delimiter filtering matches the cursor exactly

Arms that capture AST fragments (expr, ty, pat, ...) keep the parser-driven path, selected by a flag computed once per rule at definition time. An earlier variant of this PR materialized parsers for those from the flat path; it measured as a net loss on fragment-heavy crates (html5ever +3.9% at worst) and was dropped. The parser-driven input is now also built lazily per invocation, so all-flat definitions never construct a Parser at all, and all-parser definitions pay nothing for the flat machinery.

Three codegen details that each showed up as whole-crate regressions during development, now fixed and documented in the code: parse_tt_inner must stay inlined into both matchers, the flat input caches its loaded start state so rewinding for the next arm is three copies rather than a rescan, and parse_tt_flat is inlined into its single caller to drop per-arm call overhead.

Performance (ThinLTO stage2, jemalloc, instructions:u, full scenario, on top of the perf stack; base includes #82+#83+#84):

• tt-muncher check -6.05%, debug -4.11%
• clap_derive check -1.48%, debug -0.56%
• geomean -0.60% over the 10-crate x Check/Debug mbe suite
• residuals: html5ever check +0.26% (per-arm dispatch on many-arm mixed flat/parser definitions; down from +3.9% in the dropped variant), token-stream-stress check +0.29% on an 83M-instruction cell (noise class)

One caveat: several of the fixes here are inlining-shaped, and this config is plain stage2 without PGO; upstream dist builds make their own inlining choices, so these numbers want a rust-timer confirmation before any upstream attempt.

Correctness: stage2 bootstraps std and the compiler entirely through the flat matcher; full tests/ui failure set byte-identical to the pre-existing baseline (21301 passed, zero new diffs); x check compiler green standalone on this branch; a targeted smoke test covering ident/lifetime/tt captures, vis-adjacent shapes, separators, and the ty-fragment >>-splitting case produces byte-identical -Zunpretty=expanded output vs the parser-driven matcher.