BF: Seek file data to 0 before every retry in request(), not just for retryable status codes

Previously, data.seek(0) was only called inside the retry_statuses/retry_if
branch, which handles specific HTTP response codes (429, 500-504, and
zarr-specific conditions).  When a ConnectionError occurred mid-upload
(e.g., connection dropped during a multi-minute large zarr chunk transfer),
tenacity caught the exception and retried, but the file pointer was left at
whatever position the read was interrupted.

On retry, requests computed Content-Length as (file_size - current_position)
and sent only the tail of the file.  S3 received partial data whose MD5
didn't match the Content-MD5 header (computed from the full file), resulting
in a BadDigest 400 error.  This error was also not retried (not in
RETRY_STATUSES, not matched by retry_if), so the upload failed permanently.

The fix uses tenacity's before_sleep callback to seek the file data back to
position 0 before every retry, regardless of whether the retry was triggered
by ConnectionError, HTTPError, or a retryable status code.

Evidence from two independent logs (issue #1821):
- Linux: 188 level-0 zarr chunks failed (large, ~6 min upload each), 67
  level-2 chunks succeeded (small, fast upload) -- all with "succeeded after
  1 retry" + BadDigest
- Windows local disk: 187 level-0 failures after 2-8 retries, 68 level-1
  successes -- ruling out NFS/filesystem as the cause
- Both logs show "Resetting dropped connection: dandiarchive.s3.amazonaws.com"
  confirming ConnectionErrors preceded the BadDigest errors

Closes #1821

Co-Authored-By: Claude Code 2.1.81 / Claude Opus 4.6 <noreply@anthropic.com>

Author: yarikoptic

dandi/dandiapi.py

@@ -212,6 +212,14 @@ def request(
 
         lgr.debug("%s %s", method.upper(), url)
 
+        def _rewind_data(retry_state: tenacity.RetryCallState) -> None:
+            # After a failed attempt (ConnectionError mid-upload, HTTPError,
+            # etc.), the file pointer may be at an arbitrary position. Seek
+            # back to 0 so the next attempt sends the complete body.
+            # See https://github.com/dandi/dandi-cli/issues/1821
+            if data is not None and hasattr(data, "seek"):
+                data.seek(0)
+
         try:
             for i, attempt in enumerate(
                 tenacity.Retrying(
@@ -225,6 +233,7 @@ def request(
                     ),
                     stop=tenacity.stop_after_attempt(REQUEST_RETRIES),
                     reraise=True,
+                    before_sleep=_rewind_data,
                 )
             ):
                 with attempt:
@@ -249,8 +258,6 @@ def request(
                                 url,
                                 result.text,
                             )
-                            if data is not None and hasattr(data, "seek"):
-                                data.seek(0)
                         if retry_after := get_retry_after(result):
                             lgr.debug(
                                 "Sleeping for %d seconds as instructed in response "