[Performance] Rewrite the thread_safe cache.

[DarkSharpness]

1. Add a cpp test on cache efficiency and correctness
2. Refactor the thread_safe_cache.

In the new thread_safe_cache, we maintain only 1 unified mapping instead of 2. In addition, in our implementation, we acquire the unique_lock less, which may bring a significant improvement in efficiency in some cases.

Here's the result for TEST(XGrammarParallelTest, CacheEfficiency) on my machine AMD EPYC 7R32.

End to end time (preprocessing time 2s not included):

• Old cache: 1152ms
• New cache: 217ms

gcc version: 13.2.0