BUG-8403: move successor allocation to AbstractProxyTransaction

We still have a tiny race window where we do not correctly handle
reconnection, leading to an ISE splat -- this time if the final
stage of transaction completion.

The problem is that a reconnect attempt is happening while we are
waiting for the transaction purge to complete. At this point the
transaction has been completely resolved and the only remaining
task is to inform the backend that we have received all of the
state and hence it can throw it out (we will remove our state once
purge completes).

We are still allocating a live transaction in the local history
and the purge request replay does not logically close it, leading
to the splat.

To fix this, we really need to allocate a non-open tx successor,
which will not trip the successor local history. All the required
knowledge already resides in AbstractProxyHistory, except we do
not have a dedicated 'purging' state.

This makes the first step of moving the allocation caller to the
appropriate place.

Change-Id: If82957019b478f4d5132edda4d38e6bc026aa0ab
Signed-off-by: Robert Varga <robert.varga@pantheon.tech>
(cherry picked from commit cc5009b8f3ea91f64ee48cda815c6a5e73a8a1af)