Make Units set operations linear instead of quadratic

[JustinAWei]

Units.__or__, __add__, __and__ and __sub__ check each unit's tag against a generator over the other collection:

other_unit.tag not in (self_unit.tag for self_unit in self)

That rescans the whole collection for every unit, so the cost is O(n·m). This PR collects the tags into a set once, so each check is O(1). The result is the same: same units, same order (self first, then the new units from other).

Why it matters

ares-sc2 merges its remembered (snapshot) units into the live unit lists every step:

self.ai.enemy_units = self.ai.enemy_units + memory_units
self.ai.all_enemy_units = self.ai.all_enemy_units + memory_units

Against a creep-heavy Zerg, all_enemy_units reaches several thousand entries (every creep tumor seen). In a profile of our bot against such an opponent, these merges were the largest single cost in ares' per-step update, about 2 ms per call. Together with other per-step work, that pushed the average step over the AI Arena 40 ms limit.

Benchmark

A plain CPython micro-benchmark of the + logic, run in this repo's uv env:

self + other	before	after
200 + 50	0.465 ms	0.011 ms
1000 + 200	8.45 ms	0.048 ms
3000 + 300	33.5 ms	0.115 ms

Tests

• test_units gains assertions on all four operators that pin down deduplication and order.
• uv run pytest test/test_pickled_data.py test/test_expiring_dict.py: 19 passed.
• ruff check and ruff format --check are clean.

🤖 Generated with Claude Code