refactor(line): replace the nth kernel and grouped fold with 1 gather expr

[jvdd]

nth downsampling is now a single Polars gather (_nth_agg_expr) whose positions come from the column's own len(), so the same expression serves the resident select, the grouped group_by on both source kinds, and the geo line. It returns exactly what the Rust kernel returned: stride max(1, len // n_points), at most n_points rows, every row when short.

Removed:

• the every_nth Rust kernel, its Python binding, namespace method and tests
• grouped_nth_plan, the collect_batches fold and the _NTH_* constants
• _geo_line_nth_agg_expr and geo_line's plugin import

Kept nth_plan for the ungrouped scan: its two streaming passes stay flat in rows and its frame-level viewport filter prunes row groups, which the expression cannot do.

A grouped nth on a scan now gathers in memory, because Polars 1.44 streams no non-reduction aggregation. test_ooc.py marks line-grouped-nth a strict xfail. Zoomed, its memory follows the visible window and it is 2.7-3.4x faster than the fold; unzoomed it is O(rows), which was accepted.

The fold was also the site of two Polars hazards found on 16- and 32-thread machines: the streaming engine runs ahead of a slow batch consumer by about 25 MB per thread, and a sparse filter at low thread counts can return a zero-copy view that pins whole batches.

BREAKING CHANGE: pl.Expr.flexviz.every_nth no longer exists.