Sampling: Do not raise on not converged samples

See merge request !835

Better Lanczos interface

See merge request !832

re: cg: Relax energy error criterium within Newton

See merge request !833

HarmonicSKI: Make pad symmetric

See merge request !834

Operator: activate partial_insert by default

See merge request !790

FIX: add missing sqrt, save real std

See merge request !830

Implement refinement of HEALPix

See merge request !809

Do not enforce a specific dimension where unnecessary (e.g. in
`HEALPixChart`). However, do so where the computation either would fail
or would produce bogus results if invoked with fewer or more dimensions.

AbstractModel: Dynamically initialize attr

See merge request !831

SciPy does not compute the energy, so at least we should not do worse
than SciPy...

Approximates tr(log(M)) using tr(log(T)) where T is the projection of M
into the krylov subspace K(M, v) where v is a sample from the metric M
(i.E. v = v_lh + v_pr where v_lh/v_pr are samples from the
likelihood/prior metric, respectively. In addition, the projected sample
is constructed by taking v_pr projecting out the subspace K(M,v) using
its eigen-basis. This ensures that both, the prior dominated part of v
and the part already covered by tr(log(T)) is projected out.

Original author: Philipp Frank

Drop unused code blocks and refactor the geomap energy computation.

Furthermore, add a remark on the necessity of sampling antithetically.

Instead of grabbing the Hamiltonian from the global scope, require it as
input.

Unfortunately, it performs much poorer than expected.