Segfaults in Leap tests

Job #56723 failed for f894bd57.

The relevant tests die with a segfault and a moderately non-helpful backtrace:

Program received signal SIGSEGV: Segmentation fault - invalid memory reference.

Backtrace for this error:
#0  0x7fdee9a6e89d in ???
#1  0x7fdee9a6dae3 in ???
#2  0x7fdee96dafbf in ???
#3  0x7fdee9c71e70 in ???
#4  0x7fdee9c74096 in ???
#5  0x7fdee9c74335 in ???
#6  0x7fdee9c75044 in ???
#7  0x56145b904ac3 in test_rkmethod
	at /tmp/tmpdqDGXt/test_rk.f90:49
#8  0x56145b904f63 in main
	at /tmp/tmpdqDGXt/test_rk.f90:4

My guess is that compiler changes in gfortran (which recently upgraded to 8.2 on the machine shop) is exposing something that we're doing wrong. I ran with gfortran-7 to check that theory, and lo and behold, that worked around the issue for the time being. dagrt!17 (merged) together with setting this env var here (MR will follow) should get us back to passing CI for the time being.

@cory, could you take a look at (a) whether that compiler theory holds water and (b) if you can see anything problematic (e.g. using valgrind) that we might be doing?

cc @mattwala