Add IntervalFilterSpec for validity-interval pushdown

[hintse]

Summary

FilterSpec maps one output column to one source column, so it cannot express the overlap predicate needed for validity-interval / history-of-change sources, where each row is valid over [start, end] and a request date range [lo, hi] selects covering windows via start <= hi AND end >= lo.

This adds a first-class IntervalFilterSpec that pushes the correct overlap predicate.

What it does

IntervalFilterSpec(
    start_col="valid_from",
    end_col="valid_to",
    closed="both",      # "both" | "left" | "right" | "none"
    value_mapping=...,  # offsets the request bounds before pushdown (as FilterSpec does)
)

For a request range [lo, hi] extracted from the pushed predicate, it filters the source with end_col >= lo and start_col <= hi (operators adjusted for closed), mapping the bounds through value_mapping first. The overlap is resolved entirely at the source row, so the surviving rows are exactly the overlapping windows and the range predicates push through to SQL/TickStore.

Date bounds compared against a Datetime source column reuse the existing _extend_dates_to_full_datetimes widening, so intraday rows on the boundary day are not silently dropped.

Notable shared-path change

The post-combine result_lf.filter(predicate) is now restricted (via the existing restrict_expr_to_columns) to columns present in the combined output before being applied. An interval spec's request/output column (e.g. date) is supplied by another joined frame and is not a column of the interval source itself, so restricting keeps it from leaking into the source-level filter. This is a no-op for FilterSpec, whose output column combine produces.

Testing

test_pushdown_combine.py: 92 passed (no regression). Consumer-facing tests (overlap correctness, pushdown verification, closed variants, value_mapping, Datetime sources, missing columns) live downstream.

🤖 Generated with Claude Code

[github-actions]

Test Results

1 608 tests  +18   1 560 ✅ +18   50s ⏱️ ±0s
    2 suites ± 0      48 💤 ± 0 
    2 files   ± 0       0 ❌ ± 0 

Results for commit 308e717. ± Comparison against base commit 9218944.

♻️ This comment has been updated with latest results.

[ptomecek]

Nice addition — the two one-sided range predicates are exactly what pushdown can consume, and the Date-source path looks correct for all four closed modes. One correctness issue on Datetime sources with the half-open modes (details inline), plus a couple of smaller robustness notes. It'd also be good to land at least one in-repo test for IntervalFilterSpec so the module's contract is guarded here rather than only downstream.