It is technically feasible, but likely would be substantively different from what Apple did with the MP 2009-2012 models. Several points.
1. The tray will likely be large and complicated. 2009 tray was not a simple 'socket' like alternative. Neither did it plug into a commodity socket. The MP 2013 CPU was on a 'tray' also.
Neither one of those were a commodity part that Apple sold at highly affordable prices.
2. The SoC is quite likely going to serve as a "Black Hole" effect and pull more stuff onto the tray (besides the GPU). The SSD controller is in the SoC. So the SSD Modules connectors would likely get pulled back over also.
Since the GPU and Thunderbolt controllers, highly likely would also get at least a couple of TB and perhaps a HDMI socket attached to the card also. ( GPU outputs are often provisioned off the same 'card' that the GPU package is soldered to for good reasons. )
The RAM? Highly likely attached.
The System Management Controller (SMC) is relatively tightly coupled to the SoC (some tasks split between the two) so it too.
The old MP 2009 era board had the "Southbridge" on the logic board.
The SoC pulls that across onto the 'tray' also. So instead of having just the ESI/DMI link across the tray edge you now have all of that 'fan out' from the Southbridge adding to the complexity of the tray connector.
( a contributing factor to why some of the I/O like display out and at least some Thunderbolt would be
far easier just to run out the back end of the tray. )
The other offset would likely be that wouldn't get two x16 PCI-e v4 lanes run out to slots but into a two input PCI-e PLEX Switch (like on the MP 2019). Just to cut down on lanes have to traverse off the card. Run 32 lanes and then fan out on the logic board. Additionally, the PCI-e switch will serve as a 're-driver' since there are much more restrictive distance limits for PCI-e v4 (and up ... faster pragmatically gets shorter. ).
Maybe another eight other x1 PCI-e v4 off the tray to do stuff like the 10GbE , Wi-Fi , etc.
The Mac Pro 2009-2012 is a fairly tall system but ends up with only 4 slots. That is an internal space trade off. Apple probably would have to 'blow away' Slot 8 in a MP 2019 chassis to both put in a monster proprietary connector and allow enough 'tall' clearance for the large heatsink. ( instead of width of the chassis now have to consume height. )
3. All of this highly likely would not be longitudinally compatible. 2-3 years down the road Intel CPU sockets aren't compatible. That SMC controller is going to need to interact with the Power Management Controller. Generally DDR5 DIMMs don't fit in DDR3 sockets. Modularity, in and of itself, doesn't buy future computability.
Same baseline reason why SSD modules 3-5 years down the road probably won't work with a 5 year old SSD controller. (even if tried to shovel the SSD module connector to the main logic board , it only presents long term problems). The more evolutionary stick I/O try to run off the tray to drive down the costs of the tray , the more the tray will get fixed in time.
If try to run TBv4 out to ports that hand off the main logic board. Are those paths and edge PHYS logic packages going to work with TBv5 ? Probably not. Current USB4 paths going to work with future USB5 paths and PHYS? Probably not.
And when did Apple show huge interest in selling MP 2010 trays to 2009 folks? Or MP2013 trays to other 2013 folks? It is still the same company.
The 2009-2012 tray was largely a inventory cost saving measure by Apple to combine two sub-products into a single shared main logic board ( single socket and dual socket workstations sharing more designed components). It wasn't a super deep love infatuation with modularity just for modularity. Nor was it sell commodity computer component parts for next 5 years exercise. Most 2009-2010 era dual socket boards from competitors were larger ones where could fan out past the 8 DIMM slots that Apple provisioned. That tray capped it 8 because it has far less room than a conventional larger main logicboard.
Even if got a tray it extremely likely would support 10 years of updates. Most likely Apple would move from one extremely proprietary tray socket to another once got the next Mn Ultra/Extreme update. The used tray market would last for 10 years, but Apple the container on active support for 10 years? Probably not. Apple's Vintage/Obsolete support policies start the countdown when the product is superceded and withdrawn from sales. Selling 'next gen' Mn Ultra/Extreme trays (and hence new MP) would put the old one on a countdown clock that at best runs 7 years.
It is unikely that Ultra/Extreme updates will come yearly, but also unlikely that they are going to be on 3 or more year update cycle either. Probably closer to 2 years ( over 1 and less than 3 ). 2+7 isn't going to get you to 10.
