feat: make data file parts serializable

[Xuanwo]

Distributed writers need to persist a target and completed parts so another process can resume writing or assemble the final file. DataFileTarget and DataFilePart now support serde, and Lance creates staging files when writing parts. Assembly reopens and validates the referenced files before producing the final DataFile.

Managed Blob payloads are written directly into the final target's sidecar namespace and are not copied during assembly. Callers explicitly invoke target.finish(dataset) after commit to remove staging parts, or target.cleanup(dataset) after abandoning the target to remove staging parts, the final file, and Blob payloads. Both operations tolerate missing objects and can be retried after partial cleanup. Callers remain responsible for stopping users of the target, retaining the same dataset/base mapping, allocating disjoint Blob ID ranges across retries, and coordinating checkpoints with dataset GC.

This replaces the high-level caller-provided Writer and open_part flow with serializable part descriptions. Callers use Dataset::concat_data_file_parts and construct DataReplacementGroup themselves; fragment row coverage and commit fencing remain their responsibility. No separate FileFragment::write_columns_from_parts entry point is retained.

Checkpoint contents must be trusted, and callers must fence stale workers so they cannot resume writes or assembly after a target is committed. These descriptors do not prove ownership or consult retained manifests. finish retains all Blob payloads under the target, including unused retry leases; ordinary dataset GC also preserves them while the parent file is referenced. Per-Blob reachability collection is outside this change.

Follow-up to #8923.

[lance-gatekeeper]

⚠️ Gate recommendation: approve with a non-blocking risk.

The serializable descriptor design reopens and validates staging files before assembly, with coordinator fencing and fragment row coverage explicitly owned by callers. The author accepts that finish does not reclaim Blob payloads from lost or unused retry leases, so a referenced target can retain those payloads indefinitely. No further change is requested for this pull request.