If your mITX case is wide enough, you can house two CPUs. G34 example
and LGA-1366 example
. Any other board starting with "H8DGT" or "X8DTT" should have a similar form factor as well; there are a number of variants with some special features. I think it's possible to stick one of these in a Cooler Master HAF Stacker. I recommend the 915F over the 915R.
Now, as for why we don't have this today. Well, CPUs have gotten big. Intel's old server socket used to be LGA-771. That was their first land grid array socket, succeeding the pin grid array mPGA-478. Then it was replaced with LGA-1366, adding more cores and memory channels, forcing more pins for more I/O. Then that was replaced with LGA-2011, with still more cores and memory channels. And now that's being replaced with LGA-3467
! They're getting massive. AMD's socket 939 and variants (socket 940 for workstations/servers, and then everything from AM2 to AM3+) is reasonably small, but their current CPUs on that socket run extremely hot. A full Vishera die has a 95W TDP at best. Stick two of them on one board. Now you need VRMs to power 190W of CPU minimum, and two sockets take up a bit of space regardless of board. But that neglects the 1944-pin socket G34 (thankfully a land grid array like Intel, so the pins are a bit closer together) entirely with its dual-die chips, which would be even worse for any motherboard.
Technically, using a smaller socket like C32 or LGA-1366, it would be possible to have a DTX dual-socket motherboard (in other words, ATX with 2 PCIe slots instead of 7) but there isn't really a market for it. Server boards are going to go with more expansion slots for safety rather than less to save a few rack units.
Additional memory channels don't help. Each CPU needs them. ASRock cheated with their mITX X99 board by forcing it to be dual-channel. But most of the time the minimum number of memory slots is the number of channels the CPU has. For LGA-771, that's two. Easy enough to work with. For LGA-1366, that's three. A bit more difficult but still feasible. LGA-2011, that's four. Technically possible, but every mITX board with that many slots has a tiny SoC, not a massive CPU socket plus chipset plus VRMs! And for LGA-3467, six. That's just silly. For AMD, they have two channels per die, meaning the dual-die G34 CPUs are technically quad-channel. They just use two memory controllers, similar to some of Intel's biggest 15-core and 18-core dies.
And finally, there just haven't been dual-socket enthusiast boards recently because there is no real point. The Xeons needed for this, the E5-2600 series and better, are very expensive compared to the E5-1600 equivalents. Like, a 6C/12T Haswell-EP (5930K for workstations) chip costs ~$600 if it's for a 1P system and $1500 for a 2P system. That's out of budget for, well, everybody. A single 12C/24T Xeon would I think be cheaper (e.g. the E5-1686 v3, but I can't find a price for it).
To contrast the current Xeons, 5600 series (dual-socket Westmere-EP) was much cheaper, so EVGA for example could launch the SR-2 board. Since all multi-socket Xeons have locked multipliers (a few single-socket Xeons don't, primarily the E5-1600 series) and it's nearly impossible to overclock Sandy Bridge or newer via bclk, there's no real point making a dual-socket overclocking board either. You'll maybe get 100MHz more. Know what else can get that? The next tier of Xeons!
There's also AMD who I haven't really mentioned but, well, they just don't have a presence in the server market. Single-digit percentage, maybe. I think even IBM and PowerPC beats them. I'm not aware of them ever making dual-socket boards for enthusiasts outside of one time. The exception was when they pushed that in response to Intel releasing the Core2Quad (not even a true quad-core; it used two dies!) using the OG FX CPUs.