Skip invalid individuals in evaluation streams

[alcides]

Summary

• Skip individuals whose fitness evaluation raises InvalidFitnessException.
• Count invalid evaluation attempts without storing invalid fitness values.
• Make ParallelEvaluator consume input iterables lazily in bounded batches using a reusable thread pool.
• Preserve already-evaluated individuals without re-evaluating them.
• Add sequential and parallel regression tests for invalid individuals in streams.

This addresses the evaluation-stream portion of #336: evaluators can consume streams and omit invalid candidates. The existing GP operators' population-size interfaces remain in place for the follow-up target_size refactor.

Validation

• uv run --with ruff ruff check . --fix — passed.
• uv run pytest tests/core/fitness_helpers_test.py -q — 4 passed.
• Full uv run pytest -q reached 455 passing tests, then was interrupted after approximately five minutes in the existing deep-tree tests/gp/parallel_test.py; no test failure was reported.