Switching around the GPUs did reveal some interesting insights.
The good news first. Even though dmesg, lspci and even GPU-Z (in the vm) still show the same speeds as before, both GPUs now show the expected performance, indicating that the bandwidth negotiation somehow just works.
The bad news is, that the 1080 Ti does not reset correctly on slot PCIEX16_1 when it did on port _2, as well as the inconvenience that the 1080 Ti acts as the primary video output during boot as long as there is an output connected i.e. the dummy display for LG.
Conclusions:
- 1080 Ti PCIe speed shown by dmesg is not indicative of the previously bad performance
- GPU-Z really cannot be used to check PCIe bandwidth in the Windows VM
- there is a difference in behaviour between both PCIe slots when there shouldn´t be (probably a firmware issue with the motherboard)
Going forward, I will be trying a few more things:
- Increasing VFIO VGA Performance - #129 by gnif
with this I may be able to overwrite slot 2s behaviour - replacing the 1080 Ti with a 2070 I can borrow coming end of next week (maybe a newer card behaves differently)
Before that I will have to fix my 6900XTs waterblock whose flow plate somehow got dislodged 
Edit: Further testing will have to wait until I receive the replacement plexi top for the 6900XTs waterblock.