otf-pixels-codec-jpeg@0.2.0

Changed

• The README and every crate's README follow the Open Tech Foundation layout:
  a centred title and tagline, short sections, and tables instead of long
  paragraphs. Each crate's README says what it is for, when to use
  otf-pixels instead, and shows a working example (all are compiled and
  run against real files) with a table of its API. The main README adds a
  comparison with libvips/sharp, ImageMagick, Pillow and the image crate,
  with measured timings, and names ES-Runtime as the project built on Pixels.
• Every output now runs on Scheduler::global(), one process-wide pool of
  one worker per core, unless given a scheduler of its own (ADR-0016).
  Before, each output spawned a pool for itself, so a host running 40 jobs
  at once ran 40 pools on the same cores unless it knew to pass a shared
  scheduler, which the first host to embed Pixels did not. 40 concurrent
  PNG-to-JPEG thumbnails run about 13% faster. threads and
  scheduler_options still give a run a private pool, and their docs and
  EMBEDDING.md now say to leave them unset in a concurrent host. A run
  started on a pool worker gets a private pool, so it cannot wait on the
  pool it is occupying. Idle workers sleep until there is work instead of
  waking every millisecond.
• The README's "Status" section, a milestone-by-milestone progress log, is
  removed now that the v1 milestones are done; ROADMAP.md and the changelog
  keep that history.

Fixed

• Opening a WebP decoded the whole image, so asking its size or metadata
  cost a full decode (about 30 ms for a 1280x800 photo). Opening now parses
  only the container and the bitstream header, and the image decodes when
  its first row is read; metadata takes microseconds. A bitstream damaged
  past its header is now reported by the run rather than by opening.
• EMBEDDING.md recommended with_threads(1) on an output, which does not
  exist; the threading section is rewritten around the new default.
• PNG decoding was about 8x slower than it should be (216 ms for a
  1280x800 photo, now 18 ms), almost all of it in inflate, which ran some
  360x slower than zlib. It dropped consumed input after every symbol,
  shifting the rest of the buffer each time, so its cost grew with the square
  of the stream; it also cloned both Huffman tables per symbol and decoded
  codes a bit at a time. Input is now dropped only in large batches, a step
  decodes a run of symbols, codes up to 10 bits decode with one table lookup,
  and back-references copy in bulk. Inflate is within 1.4x of zlib. In the
  PNG decoder, filters are chosen once per line rather than per byte, Paeth
  reconstructs whole pixels, and 8-bit rows already in the output layout
  are copied rather than converted pixel by pixel.
• Every link in the README was broken on the otf-pixels crates.io page,
  including the embedding guide: crates.io resolved them against the crate's
  directory, not the repository root. They are now absolute GitHub URLs, and
  tsr readmes fails on a relative link in any README a crate ships.
• Only otf-pixels declared a README, so the crates.io pages of the other
  ten crates were blank. Each now has its own README.md, saying what it
  does and pointing to the otf-pixels facade, and tsr readmes (part of
  tsr ci) fails if a published crate does not ship one.