Review LOO-PL implementation for consistency/clarity/accuracy

[sgbaird]

Related reading shared by Willi, for myself to review: https://gaussianprocess.org/gpml/chapters/RW.pdf

evaluation-metrics/evaluation_metrics.py

Lines 128 to 143 in 68a2be8

	def loo_pseudo_likelihood(model, train_X, train_Y) -> float:
	"""
	Negative mean leave‑one‑out log‑likelihood (lower is better).
	Note: The GPInputWarning about matching training data is expected here
	as we're calculating LOO-PL on the training data. Not sure how to deal with the warning...
	"""
	loo_mll = LeaveOneOutPseudoLikelihood(model.likelihood, model)
	# GP output at the training inputs
	with torch.no_grad(), fast_pred_var():
	f_dist = model(train_X)
	target = train_Y.squeeze(-1)
	# average over points
	return (-loo_mll(f_dist, target).mean()).item()

[willigott]

Thanks! Relevant section is 5.4.2 (page 116 in the above pdf). Implementation can be found here.

[willigott]

The function is now - after the restructuring - here. The version in evaluation_metrics.py can be ignored. It is essentially unchanged, just had to take care of one linting issue.