Very different results moving to mace-torch==0.3.7

[WardLT]

The finetuning for MACE potentials yields very different results in MACE 0.3.7 vs 0.3.6. I'm not yet sure why

[WardLT]

It could be a change in the shape of the atomic energy dictionary? ACEsuit/mace@v0.3.6...v0.3.7#diff-2ce94c04c13013b923402cbfe1bbb205ed5a93aca0838ea63a0b95fb294b8145R160

[WardLT]

The issue seems to be only for training. Inference on models seems unaffected.

The energy shift calculation for 0.3.8 is -128 eV/atom and only 0.04 for 0.37. Let's look there.