registry: pull large blobs from mirrors that don't support byte ranges

[Outstanding2291]

Some Hugging Face mirrors (Jfrog Artifactory's /resolve/ in particular) answer
a range request with the whole body as 200 instead of 206 Partial
Content. The puller treated any non-206 as fatal "corrupt data", so a
model whose weights blob exceeds the single-stream threshold (32 MiB)
could never be pulled from such a mirror — which is the default for any
large model hosted there.

Two changes:

• A mirror that ignores Range now falls back to a single stream instead
  of failing. Ranges are only a download strategy; the blob is streamed
  once and verified by digest as before. A new Error::NoRanges signal
  replaces the misleading Corrupt error so the fallback is explicit.
• A 408 Request Timeout on that single stream is now retried with
  backoff (it was treated as terminal, since is_retryable only retried
  connect/body/timeout/5xx). A mirror or proxy that abandons the
  transfer transiently no longer kills the pull outright.

Covers both the range path (fetch_ranges) and the single-stream path
(fetch_whole): both gate retries on Error::is_retryable.

Regression tests: pulling a blob from a server that ignores Range, and
one that answers a single 408 before serving the blob.

[cobanov]

Merged, thanks @Outstanding2291. A small follow-up on main (38ae7bb) corrects the 408 test's comment: with a 40 MiB blob it retries the range request, whose 200 then falls back to a single stream.