Add openalex and format - #1

Three small additions to the benchmark harness:

1. Raise the serial batch when the longest document requires it
   iterate_many's batch must exceed the longest document, or the serial
   baseline fails with CAPACITY. The batch now grows automatically when a
   corpus demands it; the established corpora keep their 1 MiB batch, so no
   existing measurement changes.

2. Add the OpenAlex authors dataset
   A new paper dataset: the 232,330 author records updated on 2026-03-30 in
   the OpenAlex snapshot (CC0, ~6.1 GB, largest record ~1.37 MB), hosted on
   Zenodo and pinned by record and size. datasets.sh fetches it as part of
   the default corpus (reused via --corpus-from when the peer has it), and
   the query $.display_name, $.works_count is implemented in all six
   engines (simdjson, Pison, yyjson, RapidJSON, Boost.JSON, nlohmann/json).

3. Add a stream-format overhead study (comma vs newline)
   An opt-in format section measuring the same documents as
   newline-delimited and comma-delimited input, serially and on simdjson's
   two-thread pipeline — the measured cost of the comma-delimited format.