make it compatible with Coq 8.7, and test that

Turns out Coq 8.7 is 20% faster. :-) Some of this is lost to parallelism though, it seems.

8.6:

real	11m7.102s
user	25m5.797s
sys	0m17.457s

8.7:

real	9m0.417s
user	19m40.020s
sys	0m16.213s

I reported the fact that I had to change something to make it work again as https://coq.inria.fr/bugs/show_bug.cgi?id=5749.