which combination is the best ?

Alternative approach: No HA, 1x Node PROD with lightning fast PBS (dedicated x0xG NIC) all n x minute Snapshots, 1x Node WARM (empty). RT0 not zero, but (-) (-) complexity. P.S. i saw too many AUTO-vmotion failed in my lifetime.
 
This is very hardware-dependent, but if you meet the requirements for solid performance/reliability with Ceph.

We would lean towards 4 Ceph nodes + Qdevice with what little information I have.
You could co-install PBS+PVE on the fourth node or use a VM. Both have positives and negatives.
Again, it's hardware-dependent, and your hardware situation may change the recommendation.
Workload is also a factor in the decision.
 
If you have more hardware in other locations, You can sync your local PBS elsewhere to provide redundancy outside the cluster.