Inconsistent prediction scores between unbabel-comet 2.2.4 and 2.2.5

[de9uch1]

Hi. I'm developing mbrs, a library for MBR decoding, and incorporated your xCOMET-lite.

I found the inconsistent results between the versions of unbabel-comet: 2.2.4 and 2.2.5.
This is caused by this change ( Unbabel/COMET@0ccf951 ), and now different parameters are passed from when you released xCOMET-lite.
Specifically, layer_transformation argument was not passed and set to softmax automatically before, but now sparsemax is set.

What makes things even more complicated is that, because save_hyperparameters() is called in the __init__() of the XCOMETMetric class here, i.e., sparsemax, which was not actually passed before, was saved in hparams.yaml. https://github.com/NL2G/xCOMET-lite/blob/main/xcomet/hparams.yaml

Fortunately, this problem can be fixed by explicitly passing layer_transformation="softmax" in here, but I created an issue instead of a pull request because I'm not sure if this would be the change you intended.

Thank you.