Discovering the replica-set state as part of the snapshot acquisition is currently using a thread safe path that caused a performance regression under certain workloads with high concurrency.
Given that:

• We only need to know If writes are allowed (and not the precise state of the replica set node)
• For locked acquisitions, RSTL lock is taken, which guarantees serialization with stepdown
• For lock-free acquisitions, the replication term is checked optimistically to verify no stepdown occurred

We can bypass the mutex within the shard role for the `getNodeRole` helper that used on purpose
I updated the comment to warn developer to carefully use the `getNodeRole`