bug: manifest_scan_batch_size=0 can trust a stale manifest hint

Symptom

When manifest_scan_batch_size is configured as 0, manifest discovery can
silently trust a stale version_hint.json. A writer or reader may then operate
from an old manifest version and observe stale epoch, generation, or cursor
state.

Reproduction

1. Write shard manifest versions 1, 2, and 3.
2. Replace the best-effort version hint with version 1.
3. Construct the manifest reader with manifest_scan_batch_size=0.
4. Call read_latest().

The current implementation returns version 1 instead of discovering version 3.
The zero-sized 0..batch_size range schedules no HEAD requests, so the forward
scan terminates immediately.

Expected behavior

• ShardWriter::open rejects manifest_scan_batch_size=0 with a contextual
  InvalidInput error that names the parameter and value.
• The manifest reader defensively rejects zero even when constructed outside the
  writer configuration path, rather than returning stale state.
• A regression test covers versions 1 through 3 with a hint rolled back to 1.

Scope

This does not change the default batch size or the manifest file format.

Start at ShardWriter::open and the manifest reader construction path, then reproduce the rollback from hint version 1 to manifest version 3 with manifest_scan_batch_size=0 and call read_latest(). Done means both paths reject zero with contextual InvalidInput errors and a regression test covers manifest versions 1 through 3 with the stale hint.