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

[khvn26]

Contributes to #198.

• Adds an optional lazy path to Flagsmith.get_identity_flags for local evaluation. The returned Flags now 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.
• 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; set to False to opt back into the legacy eager path with the exact previous behaviour.

Why

From customer report: 420-feature local-eval env, hot loop reading one boolean flag, 400+ µs per call. Every call was re-evaluating the whole environment document even though the caller only reads a single flag.

Bench (off main, poetry run python -m benchmarks.bench)

scenario	eager	lazy single	lazy rotating	lazy all_flags()
customer-shape (0 overrides)	432 µs	1.85 µs (234×)	1.97 µs (219×)	268 µs (1.6×)
flag targeted by 10 overrides	431 µs	44 µs (10×)	2.07 µs (200×)	311 µs (1.4×)
segmenting-heavy (200 segs)	1204 µs	1.96 µs (610×)	2.13 µs	267 µs (4.5×)

.all_flags() is faster too because each lazy per-flag resolution walks only the segments that target that flag (via the reverse index), rather than every segment across every feature.

Back-compat

• Flags public surface unchanged (is_feature_enabled, get_feature_value, get_flag, all_flags, dataclass shape).
• FlagResult construction reuses the same engine helper as the bulk path — identical output shape, no reason-string drift, no hidden branching.
• Evaluation-error timing shifts from get_identity_flags to the first flag read; wrap-in-try keeps semantics for any caller that already defends against it. lazy_identity_evaluation=False restores the old timing byte-for-byte if needed.

Tests

• 8 new unit tests in tests/test_models.py (match / no-match / per-flag caching / all_flags materialisation / default-handler fallthrough / reverse-index correctness).
• 2 new integration tests in tests/test_flagsmith.py: lazy-by-default (no bulk engine call) and rollback kwarg (eager bulk call).
• 95 passing (was 87), mypy strict clean, black clean, isort clean.