fix: allow empty V1 writes with non-nullable columns

[lance-gatefixer]

Summary

• keep V1 page-statistic bounds nullable when the source column is non-nullable
• cover writing and reading an empty batch with a non-nullable string column

Root cause

The V1 statistics collector copied source-column nullability onto its min_value and max_value fields. Empty string pages legitimately have no bounds, so Arrow rejected their null statistics values while constructing the non-nullable stats struct.

Scope

This removes the panic when creating an empty V1 table. IVF_PQ index creation still requires vectors for centroid training and will continue to reject an empty dataset.

Validation

• cargo fmt --all -- --check
• cargo test -p lance-file (153 unit tests and 4 doctests)
• cargo clippy --all --tests --benches -- -D warnings

Fixes #1914

[lance-gatekeeper]

✅ Gate recommendation: approve.

The change fixes the V1 empty-batch panic at the statistics-schema boundary: min/max bounds may be absent even when the data column is non-nullable. Keeping those bounds nullable uses the existing V1 statistics grammar and preserves one statistics row per data batch.

[cursor]

Review: LGTM — approve & merge

Empty V1 pages can have null min/max even when the data column is non-nullable; copying field nullability into the stats struct correctly panics on empty batches. Making stats bounds always nullable is the right V1-stats-schema fix and is backward compatible for readers of older files. Empty non-nullable Utf8 round-trip regression is sufficient.