Derive zip member offsets from the central directory

[TomNicholas]

parse_zip_index reads every member's 30-byte local file header to find where its data starts, because ZIP permits a local header's extra field to differ in length from the central directory's copy of it. That costs one request per member — 337 for a 336-member archive — and on remote object storage it dominates the cost of virtualizing an archive.

Presume the two agree, and verify that presumption for free. In an archive written straight through, each member's data ends exactly where the next member's local header begins, and the last member's ends at the central directory. Offsets short by a stray extra field would leave a hole, so an archive that satisfies this for every member cannot be fooled by one. Archives that fail the check fall back to reading the local headers, exactly as before.

The assumption is not safe on its own, which is why it's checked rather than assumed. Measured across archives:

archive	outcome
zarr.storage.ZipStore, V2 / V3 / large members / streamed	check passes, offsets correct
336-member archive from object storage	check passes, offsets correct
zipfile with force_zip64=True	check fails → falls back (local extra field is 20 bytes longer)
Info-ZIP zip -0	check fails → falls back

No case was found where the check passes but the offsets are wrong. zip -0 matters here because the parser's own NotImplementedError for compressed archives tells users to rebuild with it, so non-ZipStore archives are explicitly in scope.

Results

Against a 336-member, 171 MB archive on remote object storage:

	before	after
parse_zip_index range requests	337	1
parse_zip_index wall time	2.53 s	0.35 s
open_virtual_dataset, one archive	7.8 s	1.34 s
throughput, 18 archives / 32 threads	0.55 files/s	3.76 files/s

The throughput number is the interesting one. Virtualizing many archives at once turned out to be request-rate bound, not latency bound — adding threads (8/16/32 identical, 64 worse) or processes (1/2/4 identical) did nothing. So eliminating requests moved the ceiling where parallelising them could not.

Tests

• test_index_does_not_read_local_headers — a 30-array archive's index costs ≤3 requests, and every member's bytes still match zipfile.read.
• test_falls_back_to_local_headers_when_offsets_disagree — force_zip64 archives read correctly, and demonstrably via the fallback rather than by luck.

Both written before the change and failing (155 and 93 requests). Full suite: 628 passed, no new mypy errors.