Old script health checks come back when upgrading a cluster

When upgrading a Nomad client in a cluster from 0.9.4 to 0.10.1, we see allocations end up with two script checks registered in Consul against them - the new one which passes, and an old one which fails. Allocations running or started on the 0.9.4 nodes in the cluster continue to function normally.

### Nomad version
All server nodes: Nomad v0.10.1 (829f9af35c77d564b3dab74454eeba9bf25e2df8)
Some client nodes: Nomad v0.10.1 (829f9af35c77d564b3dab74454eeba9bf25e2df8)
Remaining client nodes: Nomad v0.9.4 (a81aa846a45fb8248551b12616287cb57c418cd6)

### Operating system and Environment details
Debian Jessie

### Issue

Old script-based health checks come back, a problem similar to https://github.com/hashicorp/nomad/issues/5824. It seems to only happen with script checks.

### Reproduction steps

Start with a cluster running 0.9.4, with a number of jobs with health check scripts.
Upgrade the server nodes to 0.10.1
Drain a client node
Upgrade nomad on the drained client node to 0.10.1
Restart the nomad job to allow it to run on the newly upgraded node
Wait a few minutes and watch the old script health checks get re-registered against the new allocations.

I can't replicate this in development mode with a single combined server/client, but it reliably happens on our clusters.

Start by reproducing the stated upgrade sequence from Nomad 0.9.4 to 0.10.1, including draining and restarting a client, while observing script-check registration in Consul. Compare the handling of allocations and checks across old and new clients; done means old script checks are not re-registered against new allocations after the upgrade.

Written by the indexing model from the issue text.