feat: Lazy identity-flag evaluation in local-eval mode

[khvn26]

Two complementary changes targeting the local-eval latency reported in flagsmith-python-client#198 and the customer thread that motivated it.

1. Lazy identity-flag evaluation

Flagsmith.get_identity_flags returns a Flags that holds the evaluation context plus a precomputed segment-overrides reverse index, and resolves each feature on first access via the engine's already-public is_context_in_segment / get_flag_result_from_context primitives, instead of running a full bulk evaluation up-front.

The reverse index is rebuilt inside the _evaluation_context setter, so it stays in sync with environment refreshes with zero hot-path cost.

New lazy_identity_evaluation: bool = True constructor kwarg acts as a rollback switch.

2. Skip env-doc re-parse on no-op refresh

update_environment now sends a HEAD first and compares the x-flagsmith-document-updated-at response header against the value stored from the last successful fetch. When they match, the GET, JSON parse, map_environment_document_to_context, and overrides-index rebuild are all skipped — the cached evaluation context is reused. HEAD failures fall through to the existing GET path so no environment regresses if a proxy doesn't permit HEAD.

This eliminates the ~5 ms p99 GIL stall the polling thread otherwise imposes every environment_refresh_interval_seconds (default 60s) when re-parsing a 400+ feature env doc.

Bench against real customer env (Sevenrooms QA, 434 features, 64 segments, 305 rules)

Hot loop of get_identity_flags(...).is_feature_enabled(name), 20,000 iterations.

Version	mean	p50	p99	p99.9	max
flagsmith 3.8.0 (flag-engine 5.4.3)	434 µs	408 µs	726 µs	2.59 ms	25.7 ms
flagsmith 5.2.0 (flag-engine 10.0.4)	442 µs	403 µs	572 µs	4.52 ms	42.4 ms
this PR (lazy default)	2.5 µs	2.3 µs	3.4 µs	29 µs	170 µs
this PR (lazy_identity_evaluation=False)	456 µs	409 µs	1.22 ms	5.10 ms	26.2 ms

100–200× across every percentile vs any released version, on the customer's actual environment shape.

Back-compat

• Flags public surface unchanged (is_feature_enabled, get_feature_value, get_flag, all_flags).
• FlagResult produced via the same engine helper as the bulk path — identical output shape.
• lazy_identity_evaluation=False restores the old eager-eval timing byte-for-byte.
• HEAD-skip is invisible when the document changes; HEAD failures are silently swallowed and the GET path still applies.
• Two existing tests that mock engine.get_evaluation_result are pinned to lazy_identity_evaluation=False — they exercise the eager-path call shape, which still ships as the rollback.

Engine contract

Untouched. SDK only imports already-public symbols from flag_engine.segments.evaluator.

Tests

• 8 new unit tests in test_models.py (override match/no-match, per-flag caching, all_flags materialisation, fallthrough to default handler, reverse-index correctness).
• 4 new integration tests in test_flagsmith.py: lazy-by-default, rollback kwarg, HEAD-skip on unchanged docs, HEAD-failure fall-through.
• 97 passing, mypy strict clean, black/isort clean.