perf: Speed up local-evaluation hot path for large environments

[khvn26]

Fixes #198. For a 262-feature environment with local evaluation, the SDK is now ~1000x faster in the common case (no variants, no segment overrides) and 8–60% faster when the engine still has to run, depending on variant density.

Paired with Flagsmith/flagsmith-engine#296, which removes the per-feature _get_identity_key call and adds a two-key fast path for the hot get_hashed_percentage_for_object_ids case. This PR stands alone (works with the current stable engine); the engine PR unlocks additional wins on the variants path.

What changed

• Cache the output of get_environment_flags(). The evaluation context is immutable between environment refreshes, so we rebuild flags once on update instead of once per call. Invalidation is wired through a property setter on _evaluation_context, so existing direct-assignment call sites (tests, offline mode) keep working transparently.
• Short-circuit get_identity_flags() to the cached environment flag dict when the environment has no features with variants and no segments with overrides. In that case identity flags are guaranteed to equal environment flags. A fresh Flags wrapper is allocated per call around the shared flag dict so pipeline analytics still see per-identity events.
• Precompute identity.key in map_context_and_identity_data_to_context so the engine's get_enriched_context becomes a no-op on the hot path instead of performing a shallow context copy on every call.
• Pre-sort multivariate variants once at env-load time. Timsort on already-sorted input is cheap, resorting 262 lists per call isn't.
• Give Flag / DefaultFlag / BaseFlag __slots__ on Python 3.10+ (we still support 3.9) and inline the per-flag construction in Flags.from_evaluation_result, skipping a redundant helper call and truthiness check per feature.

Test changes

Two existing tests were rewritten from "assert on engine's internal call shape" to "assert on actual flag values":

• test_get_identity_flags_uses_local_environment_when_available — now verifies the returned flag values against the fixture environment instead of mocking the engine.
• test_get_identity_flags_includes_segments_in_evaluation_context → replaced by test_get_identity_flags_applies_identity_overrides, which exercises the segments-with-overrides path via the fixture's existing identity override and therefore actually verifies that segments flow into evaluation.

Behavioural contract is unchanged; both tests now fail more usefully if the SDK starts returning wrong values, and they stop being coupled to the engine's internal entry points (so e.g. engine-side refactors no longer ripple here).

Benchmark harness

New benchmarks/ directory with a local perf_counter-based harness that mirrors issue #198 (262 features, configurable multivariate density, cProfile mode). Not CI-integrated — CodSpeed would be the right place for that and would live in flag-engine. This is for fast local iteration.

Results (perf_counter, M-series Mac, 262 features)

Stock PyPI engine (10.0.3):

Scenario	main	This PR
get_environment_flags (any)	149 µs	0.07 µs
get_identity_flags vanilla	156 µs	0.6 µs
get_identity_flags 10% multivariate	156 µs	172 µs*
get_identity_flags 100% multivariate	~475 µs	370 µs*

* these paths still go through the engine fully, so their improvement is bounded by engine-side work. With the companion engine PR applied, the 10%-multivariate case drops to ~137 µs and 100% drops to ~370 µs → ~330 µs. The shortcut / caching wins above are achieved entirely in the SDK and apply today.

Test plan

• pytest tests/ (87 passing, incl. new identity-override behaviour test)
• mypy flagsmith/ clean
• benchmarks/bench.py numbers reproduce above figures
• Offline mode still populates derived state correctly (exercised by fixture tests)