fix(ci): update HF ONNX filenames and add download retries for model sync

[JerrettDavis]

The Sync Browser Models workflow hardcoded onnx/model_quantized.onnx,
which never exists upstream for Xenova/whisper-* (Whisper is a
seq2seq model split into encoder/decoder/decoder_with_past/decoder
_merged files - there is no unified model.onnx). The HEAD-then-GET
guard occasionally returned 200 on HEAD before 404ing on GET,
surfacing this as a flaky failure instead of a deterministic one.

• Replace model_quantized.onnx with decoder_model_quantized.onnx to
  match the files that actually exist on HuggingFace.
• Discover available ONNX files via the HF tree API instead of
  guessing filenames, so future upstream renames degrade to a
  skip-with-warning instead of a hard failure.
• Drop the separate HEAD probe in favor of a single retried GET
  (--retry 5 --retry-all-errors --retry-delay 3), eliminating the
  HEAD/GET consistency problem entirely.
• Align the unattended monthly schedule's default model list with
  the workflow_dispatch default (tiny/tiny.en/base/base.en), since
  small/small.en quantized ONNX files exceed GitHub's 100MB per-file
  limit and this branch has no git-lfs tracking configured - this
  was already documented in the workflow_dispatch input description
  but the schedule fallback wasn't aligned to it, which caused a
  separate "pre-receive hook declined" push failure once the 404 was
  fixed and small models were included by default again.

Co-Authored-By: Claude noreply@anthropic.com