additional features for NPSE

[gmoss13]

What does this implement/fix? Explain your changes

This introduces some additional features for score estimation named in #1226, namely:

• allow enable_transform = True for score-based potentials
• implement MAP calculation for score-based posteriors
• Implements rejection sampling for score-based posteriors to ensure prior coverage
• Allow IID observations for score-based posteriors (@manuelgloeckler has started working on this - let's discuss what's missing and merge our branches)
• implements custom converged() method for NPSE

Does this close any currently open issues?

#1226

Any relevant code examples, logs, error output, etc?

Any other comments?

• Currently, calling score_based_posterior.map() is still quite slow. We get the gradient of the log probs with respect to theta by using the score estimator, but still computing the log-probs explicitly in gradient_ascent, which is more expensive. To get around this, we save a low-accuracy ode_flow to calculate the log-probs more quickly. Ideally, we might want to write a custom gradient_ascent function for calculating the MAP for score estimators to avoid doing this altogether.
• I increased the tolerance of the test in linearGaussian_npse_test.py::test_npse_map - as far as I can tell, the reason this failed with the lower tolerance is not because of MAP calculation, but because score-based posteriors are currently slightly less accurate (at least for our test tasks).