fix(datagen): make default temporal ranges deterministic

[westonpace]

Closes #7913

Problem

The default temporal generators in lance-datagen (rand_date32, rand_date64, rand_timestamp) derived their one-year sampling range from Utc::now() at generator construction time. The RNG stream itself is deterministic, but it was mapped onto a moving wall-clock-dependent range, so the same generation parameters and RNG seed produced different arrays depending on when the generator was created. This also affected the corresponding paths through rand_type / rand_field, breaking reproducibility of saved fuzz inputs involving temporal columns.

Fix

• Anchor the default sampling range to a fixed instant: the 365 days ending at 2024-01-01T00:00:00Z (exclusive), via a shared default_temporal_range() helper. Callers that need a time-relative range can use the existing *_in_range variants, as the updated doc comments now point out.
• Add test_rng_temporal_deterministic, pinning exact expected values for rand_date32, rand_date64, and all four timestamp resolutions under the default seed. Pinned values fail on any reintroduction of wall-clock dependence (a generate-twice comparison could falsely pass when both constructions land in the same millisecond).
• Adjust the date/timestamp predicates in the test_query_date / test_query_timestamp integration tests from > DATE '2024-01-01' to > DATE '2023-07-01': with the fixed range ending at 2024-01-01 the old literals would trivially match zero rows, and a mid-range literal keeps the filters actually splitting the generated data.

Verification

• cargo test -p lance-datagen — 12 passed, including the new pinned-value test
• cargo test -p lance --features slow_tests --test integration_tests -- test_query_date test_query_timestamp — 6 passed
• cargo check --workspace --tests --benches — clean
• cargo fmt --all and cargo clippy -p lance-datagen --tests --benches -- -D warnings — clean

🤖 Generated with Claude Code

https://claude.ai/code/session_01GzLGemSfdM8i7PRi8YwJAZ

---

Generated by Claude Code

[lance-gatekeeper]

The wall-clock dependency breaks same-seed reproducibility in default temporal generation. The fixed one-year range removes that dependency across Date32, Date64, and every timestamp resolution while leaving explicit-range APIs for rolling data. Targeted generator and query tests pass, and I found no correctness, compatibility, or performance issue that should block this change.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[Xuanwo]

Thank you!