Add coverage for errors, keys, content properties and Azure and S3 operations

[tiptenbrink]

New cases, each request path verified live against Azure (flat and HNS), S3 and S3 Express:

• Error responses: Azure error codes named only in the XML body; server errors and throttling with up to five retries; S3 301 and 307 redirects.
• Response framing: answers chunked where the service chunks them, and a body that ends before its Content-Length.
• Unusual keys: ?, #, %, +, a space, non-ASCII text, // and a trailing /, on write, HEAD, delete, stage and commit.
• Write options: Content-Type, Content-Encoding (stored gzip), Content-Language, Content-Disposition and Cache-Control, read back by GET and HEAD, and If-Modified-Since and If-Unmodified-Since.
• Operations: copies, batch deletes, containers and buckets, Azure leases (including a conditional write or delete to a leased blob, which Azure refuses with 412 LeaseIdMissing although the condition holds), tags, access tiers, append blobs and page blobs, and S3 tagging and storage classes.

The grader gains what these cases need: optional trailing exchanges (retries, a resumed download), chunked and cut-short responses, and a response written in one piece. Two grading changes:

• A decline of a case whose parameter the service ignores passes only when the adapter's unsupported result names that parameter, as a refusal must. Before, any unsupported passed those cases.
• An entry of the expected-unsupported list may be a pattern, where * matches any run of characters, such as operations/s3/*. An entry may also name a result field, such as field:/value/content_md5_base64: it covers the cases that the adapter left unsupported only by declaring such fields unsupported, which a report now lists as unsupported_fields. Every case an entry covers must be unsupported, and a field entry that covers no case of a whole-suite run is a mismatch. --record-unsupported keeps patterns and field entries and lists only the cases they miss.

The example-adapter step now fails when the grader fails; piping into tail hid that. The example adapter pins borink-org/object-storage#34, so merge that first, with a merge commit.

[tiptenbrink]

Well, I did ask before merging the previous PR whether we wouldn't have a lot of churn afterwards, turns out that was wrong.