Window function perf: Cache nth_value valid indices

[Dandandan]

Which issue does this PR close?

• N/A.

Rationale for this change

Cache valid row positions for nth_value window evaluation with IGNORE NULLS to avoid rescanning validity bitmaps for each frame.

What changes are included in this PR?

• Add a lazy per-evaluator valid-index cache for nth_value IGNORE NULLS evaluation.
• Reuse cached non-null row positions across frame evaluations.
• Add regression tests for cached ranges including backward and empty-valid-row ranges.

Are these changes tested?

• cargo fmt --all
• cargo clippy --all-targets --all-features -- -D warnings
• cargo test -p datafusion-functions-window nth_value --lib
• cargo bench -p datafusion-functions-window --bench nth_value -- nth_value_ignore_nulls/last_value_expanding/50% --sample-size 10 --warm-up-time 1 --measurement-time 2

Benchmark median: 260.09 us. Original baseline: 14.239 ms.

Are there any user-facing changes?

No API or behavior changes; performance improvement only.