replication: prevent gc from deleting xlogs that haven't been fetched

Currently, nothing prevents the garbage collector from deleting xlogs
that haven't been fetched by a replica yet. To fix that, let's keep
track of the oldest checkpoint needed by each replica. The replica
checkpoint is first set on initial join - at that time it points to the
snapshot the replica was bootstrapped from. After that it gets updated
on xlog relay, by tx_status_update(). Since the latter may invoke
garbage collection now (by unreferencing the previous checkpoint) and
gc_run() may yield on coeio_unlink(), we have to use "tx" pipe instead
of "tx_prio" for communication between relay and tx cords.

Closes #748