Performance regression: v5 is 2–7× slower than v3.8.0 for local evaluation

Benchmarked `flagsmith` 5.0.1 vs 3.8.0 against the same environment (edge API, `enable_local_evaluation=True`, 262 features, single-threaded, after 200-iter warmup).

### p50 per call

| Operation | v3.8.0 | v5.0.1 | Regression |
|---|---|---|---|
| `get_identity_flags` | 0.39 ms | 0.82 ms | **2.1×** |
| `get_environment_flags` | 0.11 ms | 0.79 ms | **7.4×** |
| `is_feature_enabled` (cached `Flags`) | 0.08 µs | 0.08 µs | — |

Throughput drops from ~2,570 → ~1,215 identity evals/sec per core. Latency variance also grew (stdev ~10 µs → ~220 µs).

### Repro

```python
from time import perf_counter
from flagsmith import Flagsmith

client = Flagsmith(
environment_key="",
api_url="https://edge.api.flagsmith.com/api/v1/",
enable_local_evaluation=True,
environment_refresh_interval_seconds=3600,
)
client.update_environment()

# warmup
for _ in range(200):
client.get_identity_flags(identifier="anonymous", traits={"venue_id": "12345"})

for n in (100, 1_000, 10_000):
t0 = perf_counter()
for _ in range(n):
client.get_identity_flags(identifier="anonymous", traits={"venue_id": "12345"})
elapsed = perf_counter() - t0
print(f"{n:>6} iters: {elapsed*1000:8.1f} ms ({elapsed/n*1e6:7.2f} µs/call)")
```

Environment: flag-engine 10.0.1, Python 3.12, macOS.