Fix memory blowup in GAM.gridsearch over large lambda grids

[RogerPR]

Summary

Closses issue: #242

GAM.gridsearch previously kept every fitted candidate model in memory
for the whole duration of the search, even when the caller did not need
them. For large lam grids this caused memory usage to grow linearly
with the grid size and could exhaust available RAM.

This PR keeps only the models that are actually needed:

• best_model — to copy back into self when keep_best=True.
• last_model — used as the warm-start coefficients for the next fit.
• The full models list is only populated when return_scores=True,
  since that branch returns it to the caller.

In addition, fresh candidate models are now instantiated via
self.__class__(**self.get_params()) instead of deepcopy(self).
deepcopy was eagerly copying any large fitted state attached to
self, contributing to peak memory.

The public API is unchanged: gridsearch returns self by default and
the OrderedDict[model -> score] when return_scores=True, exactly as
before.

Changes

• pygam/pygam.py — GAM.gridsearch: only retain models when
  return_scores=True; track last_model separately for warm-start;
  build candidates from ModelClass(**base_params) instead of
  deepcopy(self).
• pygam/tests/test_memory_leak_gridsearch.py — new regression test
  that runs a 100-point gridsearch and asserts peak RSS stays bounded
  and end-of-run RSS does not grow beyond a tolerance.
• pyproject.toml — adds psutil to the [dev] extras (used by the
  new test).

Test plan

• pytest — 163 passed, 1 skipped (full suite green locally).
• pre-commit run --files <changed files> — all hooks pass
  (ruff-format, ruff, trailing-whitespace, end-of-file-fixer).
• Existing return_scores=True callers (covered by
  test_gridsearch_returns_scores, test_gridsearch_keep_best,
  test_no_cartesian_product, etc.) still pass — the
  OrderedDict[model -> score] return contract is preserved.