fix: make location spacing order-independent

[ChrisW09]

Fixes #10.

Problem

BaseLocationSelector._enforce_spacing tests each candidate with
point - spaced[-1] >= min_distance, which is only meaningful for an ascending list.
LightGBMLocationSelector._ordered_candidates returns candidates in gain-descending
order — the base-class docstring at selectors.py:121-126 states that is the contract — so
any candidate positioned below the previous keeper yielded a negative difference and was
silently dropped.

Two consequences, both silent:

1. The candidate set collapsed to a small cluster. On x ~ U(0, 10), y = sin(x),
   184 candidates became 8, all inside [6.27, 9.82].
2. Because the survivors emerged in ascending order, _trim_over_max's points[:max_count]
   then kept the numerically lowest thresholds rather than the highest-gain ones,
   discarding the gain ranking that is the entire reason to choose lightgbm over cart.

Fix

Compare each candidate against every location kept so far rather than only the last one,
making the filter independent of arrival order.

For already-ascending input the two tests are equivalent — spaced[-1] is the nearest kept
location — so the single-tree path is bit-for-bit unchanged. The gain-ordered path now keeps
its ordering, which also makes the existing points[:max_count] trim correct without
touching it.

Result

	before	after
candidates surviving spacing	8	65
select(min=3, max=8)	[6.27 … 9.82]	[2.84 … 9.14]

PLE thresholds through the Preprocessor (feature scaled to [-1, 1]):

before: lightgbm [0.26 0.29 0.32 0.82 0.87 0.9  0.94]   <- lower 60% = one bin
after:  lightgbm [-0.43 -0.4 0.18 0.2 0.26 0.29 0.32]
        cart     [-0.3  0.15 0.26 0.38 0.73 0.82 0.91]

The remaining clustering in the lightgbm row is now genuine gain ranking (the steepest
parts of sin), not an artifact of the filter.

Tests

Five added to tests/test_location_selectors.py:

• test_enforce_spacing_is_order_independent — ascending and shuffled input give the same set
• test_enforce_spacing_still_drops_close_neighbours — the spacing guarantee still holds
• test_enforce_spacing_matches_legacy_result_on_ascending_input — pins the no-change
  guarantee for the single-tree path against a literal copy of the old loop
• test_lightgbm_locations_cover_the_feature_range — regression test for the reported bug
• test_lightgbm_keeps_the_highest_gain_locations — gain order survives the filter

Full suite: 485 passed, 9 xfailed. ruff check clean on both changed files; pyright
unchanged at its pre-existing 71 errors.

🤖 Generated with Claude Code