mbe: literal-run fast path in the matcher; one-entry marker cache [html5ever -0.3%]

[xmakro]

Follow-up to #57, stacked on #83. Two independent fast paths in the mbe hot loops.

Matcher literal-run fast path. Matching stepped every literal token of an arm through the full queue machinery: a parse_tt_inner round, a queue swap, and a parser bump per token. When there is a single live matcher position and the tracker does not observe locations, a run of literal-token (and delimiter-marker) locations is now matched directly against the input:

• once at rule start, so arms that fail on a leading literal (the common case when a macro tries its rules in order) never enter the queue machinery at all, and
• again whenever a matched token leaves exactly one next position.

Failure values and token positions are identical to the queue path, matching stops at input Eof so the distinct missing-tokens diagnostic still comes from the queue path, and tracking invocations (the diagnostics retry pass) are gated out via a new Tracker::NEEDS_TRACKING const, so they observe every location exactly as before.

One-entry marker cache. Marking spans during transcription hashed the syntax context into the marker cache for every token. A one-entry cache in front of the map now hits on a plain compare, since virtually all tokens in a macro body share one syntax context.

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

• html5ever check -0.27%, debug -0.21%; clap_derive check -0.13%; deep-vector -0.13%; serde/syn -0.06 to -0.07%
• tt-muncher -0.17% (the matcher fast path is placement-sensitive: a per-token variant reached html5ever -1.06% but cost tt-muncher +1.95%; the shipped placement takes the safe side, and a per-matcher precomputed gate could reclaim the rest as future work)
• only counter-move: token-stream-stress debug +0.38% on a 150M-instruction stress cell (+568K instructions absolute, consistent across matcher-design variants, likely code layout)

Correctness: stage2 green; x check compiler green standalone on this branch; full tests/ui failure set byte-identical to the pre-existing baseline (21301 passed, zero new diffs), which includes the macro diagnostics tests exercising the tracking path.