Support GPU reallocation during HA VM failover (#270)

* Support GPU reallocation during HA VM failover

Previously HA failover refused VMs with attached GPUs. Derive the
VM's GPU requirements from its current allocation, factor them into
target host selection, and atomically move the GPU allocation to the
chosen host as part of the failover.

* Address review feedback on GPU failover reallocation

Move GPU reallocation after the VM state update so a failure can be
rolled back consistently, and revert the GPU allocation if either the
reallocation or the subsequent VM start fails. Also derive vendor/model
from the first allocated GPU that has them set, instead of assuming
index 0 is populated.