Keep the caller's parameters and the open-time pin in S3File #1334

Part of #1327 (defects 2 to 7 of the audit report).

Summary

S3File loses or overrides what the caller gave open() and what the open-time lookup resolved, in six places. Each was reproduced offline with a fake client at 766d79d.

Defects

1. A read-mode open() replaces the caller's IfMatch. kwargs.update({"IfMatch": etag}) (s3.py:2787-2788) overwrites it with the current ETag: fs.open(p, "rb", IfMatch='"stale"').read() returns the current object instead of failing with 412.
2. A reopened file reads another object with the old size. S3File.__init__ calls AbstractBufferedFile.__init__ without the request parameters (s3.py:2934-2943), so self.kwargs is {} and __reduce__/fsspec.spec.reopen drop RequestPayer and SSECustomer*. The reopened file pins the ETag current at reopen but keeps the size of the original (s3.py:2790-2791): after a 4-byte object was replaced by an 8-byte one, the unpickled file returned the first 4 bytes of the new object without an error.
3. S3File.metadata(), getxattr(), url() and setxattr() drop the file's lookup parameters (s3.py:3225-3266): for a file opened with SSE-C, the open-time HeadObject carries SSECustomer*, the one of metadata() only Bucket and Key.
4. An append's part copies are not pinned to the object looked up at open. _initiate_upload (s3.py:3083-3097) copies from S3Path(bucket, key) without the version or CopySourceIfMatch; above 5 GiB it re-runs an unpinned info() for the size, while AioS3StreamedFile uses the open-time size (s3_async.py:1526-1529). The small-object path reads the old content with a cat_file that is not conditioned on the open-time ETag (s3.py:2795-2804). An overwrite between open and flush makes the result combine the new object's bytes with the old object's metadata, or mix two objects across part copies.
5. ab and xb on a key that is only a prefix. With no object at dir and objects under dir/, wb writes, ab raises FileNotFoundError (it runs cat_file on the directory entry), and xb raises FileExistsError although no object exists (s3.py:2795-2809). Read mode already treats a directory entry as no object (s3.py:2775-2777).
6. commit() on a read-mode file raises AttributeError: 'NoneType' object has no attribute 'multipart' (s3.py:3180-3181); discard() is a documented no-op for reading.

Fix direction

Keep the resolved target (ETag, version, size, lookup parameters) on the file as the single source of the pin: respect or verify a caller's IfMatch, carry the per-open parameters and the pin through __reduce__ (deciding how SSE-C keys are pickled), route the file's metadata methods through the target's lookup parameters, plan append copies from the pinned version or ETag and the open-time size in both file classes, treat a prefix-only key as no object in ab/xb, and make commit() a no-op in read mode.