v0.7.34

Changed

• The dashboard's charts are drawn by Chart.js 4.6.1 (was 4.4.4), and the
  dashboard is built with Vite 8.3.5 and vite-plugin-singlefile 2.3.4. A real
  run's charts were compared between the two Chart.js versions and look the
  same.

Internal — which test logs are real

• Every test log now says whether it is a recording. tests/fixtures/logs/
  has a README listing each log as a recorded run, with the script that
  produced it, or a hand-written format sample, with the layout it imitates
  and whether a library really prints it; a test fails on a log in neither
  list. Checking them found that the pytorch_lightning_* samples are not
  Lightning's layout (they are a hand-written loop's tqdm bars) and that
  yolo_100ep and its siblings match no Ultralytics version. No parser code
  exists only for an invented layout.
• New and re-made recordings, each with its script: a hand-written PyTorch
  loop with tqdm bars, and Accelerate and fastai runs whose 0.5.16 captures had
  no script.
• Corrected: 0.7.33 said the extension replays 3,876 recorded parser calls.
  The test feeding real logs split every progress-bar redraw into a line of
  its own, which the pipeline never does; read as the pipeline reads them,
  there are about 2,900, and the extension still matches Python on all of
  them.
• CI no longer waits for hours on a slow package mirror: the browser job's
  system-package install is limited and retried.