[XLA:GPU] remove some redundant mutex lock in fusion analysis cache invalidation

[Cjkkkk]

[Cjkkkk]

Since we're changing the thread-compatibility expectations of the class, could you please update the comment in fusion_analysis_cache.h? To say that Get can be called concurrently, but Invalidate and Clear shouldn't.

Sure, I dont see them getting called concurrently anyway so i think we should just remove it.