Agent can become stuck in (re-)registering state during upgrades

Currently, an agent will not be erased from the set of currently (re-)registering agents if

• it tries to (re-)register with a malformed version string
• it tries to (re-)register with a version smaller than the minimum supported version
• it tries to (re-)register with a domain when the master has no domain configured
• the operator marks the slave as gone while the (re-)registration is ongoing

Afterwards, all further (re-)registration attempts with the same agent id will be discarded, because the master still thinks that the original (re-)registration is ongoing.

Since most realistic way to encounter this issue would be during cluster upgrades, and it will fix itself with a master restart, it is unlikely to be reported externally.

Review: https://reviews.apache.org/r/64506