Please reenable the WDDM policy. I’m working with RDSH for around 20 years now and never disabled this policy.
For the GPU channel count: This is a pure GPU hardware topic and has nothing to do with PCI lanes or other server hardware.
See here for example:
This is a known issue for vGPU but applies for sure the same way for RDSH as each user on RDSH requires GPU channels and especially W2022 requires a massive amount of these channels.
In addition there is an architectural change in the channel handle between Ampere and pre-Ampere generations because of SRIOV. So this could explain why it worked fine in the past with older GPUs but could now cause an issue.
At this point this is just an assumption that you may hit the channel count limit. Without enterprise support and the option to work with our engineering it is pretty hard to prove if this would be the cause of your issue.