Revert RATIS-2099 due to its performance regression

This commit creates a significant extra cost in the critical path (which is run sequentially) of Ratis appendTransaction. 

This seems to be a premature optimization. One or two instances of TermIndex per request are basically nothing (unless we create hundreds/thousands of them per request).   Short-lived POJO like this are the best to be dealt with by java GC/heap.

More details are the parent Jira RATIS-2129.