allow sampling of posterior outside of prior bounds

[rdgao]

currently, sampling from posterior rejects samples outside of the prior bounds (by design, due to leakage), but it may be useful to optionally allow sampling outside of the bounds with a flag in posterior.sample() or as an additional parameter in rejection_sampling_parameters

as discussed with @michaeldeistler

[janfb]

idea: when doing the entire SBI in transformed space, then outside posterior sampling will not be an issue anymore. we would do this by default, and when the user sets it to False we warn and samples might be outside of the prior.

[michaeldeistler]

Removing the hackathlon label because we will deal with this after porting to flowtorch

[janfb]

fix: sample from the density estimator directly and set your own rejection bounds.