perf: precompute field.relation_type and slim RelationsManager init

[collerek]

Summary

Three micro-optimizations on the per-Model.__init__ relation-bookkeeping path. Stacked on top of #1650 (the __getattribute__ cleanup).

1. Precompute field.relation_type in BaseField.__init__. The branch on is_multi / is_through / virtual / is_relation runs once at field-declaration time. RelationsManager._add_relation then reads field.relation_type directly instead of calling _get_relation_type(field) per relation per Model.__init__. Saves ~100 k method calls per all_with_related scenario; _get_relation_type deleted. RelationType is lazy-imported via a module-level _relation_type_cls() helper to break the circular ormar.relations → relation_manager → utils → foreign_key → base chain.
2. Drop RelationsManager._related_fields / _related_names. __contains__ now checks the _relations dict (already keyed by field name). Removes a per-init list-comp + two STORE_ATTR ops.
3. _add_relation reads field.relation_type directly. An assert narrows the Optional for mypy and documents the invariant (only relation/through fields reach this path).

Behavior unchanged.

Benchmark deltas (median, pytest-benchmark, --warmup, 10+ rounds)

Test	Before	After	Δ
test_initializing_models[250]	2 957 µs	2 726 µs	−7.8 %
test_get_all[1000]	17.29 ms	16.06 ms	−7.1 %
test_initializing_models_with_related_models[10]	308 µs	291 µs	−5.7 %
test_initializing_models[500]	5 462 µs	5 371 µs	−1.7 %
Single-row get_one / first	—	—	I/O-dominated, within noise

cProfile (all_with_related scenario): 5.481 s → 5.328 s = −2.8 %. _get_relation_type no longer in the trace.

Test plan

• make pre-commit — clean (mypy strict)
• make coverage — 628 passed, 25 skipped, 100.00 % coverage
• cProfile on all_with_related confirms _get_relation_type no longer in top-25 by tottime
• All targeted tests green: tests/test_relations/, tests/test_queries/test_queryproxy_on_m2m_models.py, test_aggr_functions.py, test_reverse_fk_queryset.py

Notes

• This PR's base is perf: replace RelationProxy.__getattribute__ with explicit count/clear #1650 (perf/relationproxy-getattribute) — once that merges into check-optimisations-and-rs, the diff here will rebase to just the three files shown.
• _resolve_relation_type was simplified during review from function-attribute caching to a module-level _RELATION_TYPE variable + global (idiomatic, no # type: ignore needed).
• An earlier _to_remove sentinel change (None instead of always-empty set()) was reverted during review — the saving (~7 µs / scenario) didn't justify the added Optional type + assert + conditional checks.

🤖 Generated with Claude Code