I am also on dx 12 on a nvidia card, but as you said yourself you can’t rely on dx 12, on some cards a few features will work, on some they won’t, and on some they might work sketchy 
Sounds like a big nightmare since you have to write fallbacks, and probably fallbacks for fallbacks.
To the specific async compute, i know that every one speaks about, but when it comes to useful implementations i see only gpu side culling as a really performance winning topic. Currently there are only synthetic benchmarks out there designed to show the superior drawcall performance, but in most of those scene you could come away with way less drawcalls if you used some dynamic batching. (Note, GL4.5 supports already every single feature mentioned up to this point)
The other main problem i see comming is the low level access. Considering how the big players mess up the games using a high level abstraction layer i forsee a hard time for the average gamer. The low level stuff might work good on consoles where you have a single hardware environment to consider. But on PC, with gpus VRAM ranges from 1.5GB over to 2, 2.5, 3, 3.7 (pun intended), 4, 6, 8, and even 12 GB you have to optimize your memory strategies / streaming technique for every single capacity. (If you want to get the best out of each card of course)
But i might have seen already to many highly discussed topics showing the future performance but only very few of them made it to deal changing features you can actually use in games.
And then, JME is open source. Sources for dx 12 are available, so go ahead. Nothing stops you from adding support for it. At least most of the generators are able to generate the jave bindings for dx 11. Have never tried dx 12