perf: cherry-pick PR #172 性能优化（不含 orjson）

- tiktoken 编码器改为 @lru_cache 全局缓存，避免多实例重复初始化
- 前缀匹配按长度排序，避免短前缀抢先匹配
- AuthService 更新 last_used_at 时临时关闭 expire_on_commit，减少额外 SELECT
- UsageService.record_usage_batch 改用 bulk_insert_mappings 批量插入
- 已有记录查询增加 selectinload 预加载，避免 N+1
- Gemini normalizer 长行格式化

Co-Authored-By: AAEE86 <33052466+AAEE86@users.noreply.github.com>

tests/unit/test_tiktoken_counter_cache.py

@@ -0,0 +1,41 @@
+import pytest
+
+
+class _DummyEncoder:
+    def encode(self, text: str) -> list[int]:
+        return [0] * len(text)
+
+
+class _DummyTiktoken:
+    def __init__(self, calls: dict[str, int]) -> None:
+        self._calls = calls
+
+    def get_encoding(self, _name: str) -> _DummyEncoder:
+        self._calls["get_encoding"] += 1
+        return _DummyEncoder()
+
+    def encoding_for_model(self, _model: str) -> _DummyEncoder:
+        self._calls["encoding_for_model"] += 1
+        return _DummyEncoder()
+
+
+@pytest.mark.asyncio
+async def test_encoder_is_globally_cached(monkeypatch: pytest.MonkeyPatch) -> None:
+    import src.plugins.token.tiktoken_counter as tc
+
+    # 清理缓存，避免受其他测试影响
+    tc._get_encoder_cached.cache_clear()
+
+    calls = {"get_encoding": 0, "encoding_for_model": 0}
+
+    monkeypatch.setattr(tc, "TIKTOKEN_AVAILABLE", True)
+    monkeypatch.setattr(tc, "tiktoken", _DummyTiktoken(calls))
+
+    p1 = tc.TiktokenCounterPlugin()
+    p2 = tc.TiktokenCounterPlugin()
+
+    # 两个实例对同一 model 请求编码器，底层 get_encoding 应只触发一次
+    await p1.count_tokens("hi", model="gpt-4")
+    await p2.count_tokens("hi", model="gpt-4")
+
+    assert calls["get_encoding"] == 1