KAFKA-20743: Restrict sentinel leader and state epoch in share coord.

[smjn]

• currently in ShareCoordinatorShard the writeState RPC allows negative
  leader and state epoch values, readState allows negative
  leaderEpoch and initState allows negative stateEpoch.
• though these epoch's could never be injected by the client and are
  server controlled, still we need to harden them.
• the initState call always has the correct stateEpoch (viz groupEpoch)
  by virtue of
  being called from the group coordinator (both in initializing and
  altering offsets).
• readState is called from
  SharePartition and will always be preceded by initializeState call,
  which sets the stateEpoch (non negative). Furthermore, SharePartition
  also creates new leaderEpoch per leadership change.
• writeState will always be preceded by readState which will have
  correctly set the leaderEpoch by virtue of SharePartition. And since
  initializeState correctly sets stateEpoch - writes should contain valid
  values for both the epochs.
• in light of this ShareCoordinatorShard has been tightened.
• new tests have been added to verify the change.

Reviewers: Andrew Schofield aschofield@confluent.io, Apoorv Mittal
apoorvmittal10@gmail.com

[AndrewJSchofield]

Thanks for the PR. Looks good to me.