VSR: Fix liveness for io_depth_write overflow when replicas <= 2

Where `config.pipelining_max` exceeds `config.io_depth_write` it's
possible for a client request to be unable to acquire a write IOP if we
have maxed out our IO depth.

This can lead to deadlock for a cluster of one or two, since there is
no other way for the leader to repair the dirty op because no other
replica has it.

The fix is for `on_prepare_timeout()` to retry the prepare.

Reported-by: @ThreeFx

Fixes: tigerbeetle#5

src/simulator.zig

@@ -129,19 +129,16 @@ pub fn main() !void {
         seed,
         replica_count,
         client_count,
-
         request_probability,
         idle_on_probability,
         idle_off_probability,
-
         cluster.options.network_options.packet_simulator_options.one_way_delay_mean,
         cluster.options.network_options.packet_simulator_options.one_way_delay_min,
         cluster.options.network_options.packet_simulator_options.packet_loss_probability,
         cluster.options.network_options.packet_simulator_options.path_maximum_capacity,
         cluster.options.network_options.packet_simulator_options.path_clog_duration_mean,
         cluster.options.network_options.packet_simulator_options.path_clog_probability,
         cluster.options.network_options.packet_simulator_options.packet_replay_probability,
-
         cluster.options.storage_options.read_latency_min,
         cluster.options.storage_options.read_latency_mean,
         cluster.options.storage_options.write_latency_min,