WIP: Ignore JetFitter NAN in frozen derivaton tests

Unfortunately a number of inputs to the flavor taggers are sometimes NAN. This is fine from a computing and physics point of view: it's reproducible and the downstream code handles these inputs fine. The place where it's not fine is in the frozen derivation tests: our diff script recognizes two files with identical branches as different because NAN != NAN according to IEEE standards.

Since this isn't actually a difference from the regression testing point of view the idea solution would be to make the acmd diff-root script respect NAN == NAN. But alas, there's no maintainer for acmd diff-root so I'll do the next best thing and tell the frozen derivation test to ignore some of the branches where this problem exists.

This isn't super pretty. It pokes a hole in the whole frozen derivation test: if these branches do change one might imagine we'd have no way to know. Lucky for us, they are all inputs to the DL1 and DL1r neural networks, and the nice thing about a neural network is that it's very hard to change an input without changing the output at the same time.