Just because it only has x1 lanes doesn't mean you can't put an x16 card in it it just means it'll only run at x1 when you do. Same story as why the x4 NVMe drive worked when attached to the x1 slot.
As far as the physical you can either get a riser or just Dremel the plastic back stop off (though you'd still need to figure out physical support for a larger device). Many short PCIe slots come with an open back out of the box just for this purpose, I'm surprised the Pi (for all it's hacking purpose) doesn't. Power is also going to be another thing, PCIe is supposed to provide up to 75 Watts. There are solutions for this as well, particularly if you went with a "bitcoin mining" type riser in the earlier step.
Is there any way for two or more things to share the one lane? Like two or more of: SATA, Coral [0, 1], second Ethernet, USB3? I'm guessing no, but I'm not real knowledgeable about this stuff. I think maybe the best you can do is dedicate it to USB3 and get everything you want in USB3 adapter form, just as you would with the non-compute module form.
Sure, PCIe is packet based so PCIe switches are a thing. Just be aware this is a single PCIe 2.0 lane, it's already overburdened when you added the first USB 3 port.
For the coral it's probably cheaper and just as good a fit to put a USB 2 hub chip on and wire up the B+M m.2 version through that.
Hmm, interesting, thanks, and I see now there are PCIe expanders based on that idea. [1] I think the CM-4 + reference IO card + expander + PCIe versions of stuff I'd want are getting more expensive than makes sense (in comparison to say the ODROID-H2+ that has more built in and I'm finding hard to beat lately) but maybe someone will make a single IO board that's more appropriate for a NAS or NVR. This hackaday article [2] suggests the CM-4 would be good for that.
I looked up those bandwidth numbers:
USB 2 = 480 Mbit/s
PCIe 2.0 = 4000 Mbit/s
USB 3 = 5000 Mbit/s
I keep looking for lower-cost machines for a fully-functional open-source NVR. I don't think I'd be likely to use more than 4000 Mbit/s between a couple SATA HDDs, a second Ethernet port, and a Coral or two. But the Coral can use more than 480 Mbit/s by itself. The USB accelerator product page [3] says "* Compatible with USB 2.0 but inferencing speed is slower."
In the linked video[1], Raspberry Pi foundation show an extension cable being used to move a card away from the IO board. They indicate that they expect most manufacturers to alter the design of the IO board to more closely match the product requirements, and not leave an NVME drive, or PCI express card poking up.
I heard on the grapevine though that the PCIe aperture space was tiny for the rpi4 SoC and there simply wasn't enough physical memory space to map in the BARs of a typical GPU.
As far as the physical you can either get a riser or just Dremel the plastic back stop off (though you'd still need to figure out physical support for a larger device). Many short PCIe slots come with an open back out of the box just for this purpose, I'm surprised the Pi (for all it's hacking purpose) doesn't. Power is also going to be another thing, PCIe is supposed to provide up to 75 Watts. There are solutions for this as well, particularly if you went with a "bitcoin mining" type riser in the earlier step.
Plenty of people have done GPUs on https://www.pine64.org/rockpro64/