fix: avoid inefficient collect/count at interrogation time

[rich-iannone]

In Pointblank 0.25.0, per-row validation checks on Polars LazyFrames became 8–11× slower than 0.24.0, with the worst slowdown occurring exactly when checks pass:

Check	0.24.0	0.25.0	Slowdown
col_vals_not_null	0.007s	0.077s	11×
col_vals_le	0.004s	0.031s	8×

In 0.24.0, each LazyFrame was collected once before per-row checks ran, so the results table was eager. In 0.25.0 that eager collect was removed to keep the pipeline lazy, but the code that tallies results still issued three (or four, for col_vals_in_set) separate .collect() calls on the lazy results table: one for the passed count, one for the failed count, one for nulls, and one for the total row count. Each .collect() re-executed the entire lazy plan (including the per-row predicate) from scratch.

To address this, I added the _count_validation_units() util function. It's a single Narwhals aggregation that computes the row count, pass count, fail count, and null count in one pass (one .collect()). The result-tallying block in validate.py now calls it once instead of scanning the data three to four times. Booleans are cast to Int32 before summing for PySpark compatibility, and the sums naturally exclude nulls so pass/fail semantics are unchanged.

On the 2M-row reproduction from the issue:

Check	Before	After
col_vals_not_null	0.077s	0.008s
col_vals_le	0.031s	0.009s

Performance is now back to 0.24.0 levels.

Fixes: #405