perf: cache widths, reuse buffers

Text docs keep their width, so fits no longer measures the same text
again and again, and tracks trailing whitespace only when a doc has a
trim. fits reuses one stack, texts that are part of the source are
borrowed instead of copied, and the arena and output reserve room up
front. Pages on one thread: 111 to 82 ms.

Adds the CLI, bench/run.ts, and its markup_fmt and @lukeed/html
helpers.