replace by fallback on train_dataloader

but keep it in AbstractHybridProblem  to allow returning an automatic train_dataloader

and add get_hybridproblem_HVIApproximation

to allow separating parameter and scale outputs of ML model

need to change ϕqc.logσ2_ζM_offsets from nested vector plain vector and reconstruction of inserting zeros
Interesting session with chatbot to implement performant insert_zeros.

to allow efficient indexing with a vector to insert zeros.

to allow efficient indexing with a vector to insert zeros.

…ence.jl into dev

to avoid repeated computation

with ML-predicted uncertainty scaling factor

need to provide priors_M, penalties, and logdetjac (entropy) components per site.

This required refactoring of StackedArray Bijectors and Stacked, which cannot sum lotdetjac any more.
New function with_logabsdet_jacobians (note the s)
It assume that the logjac output of a bijector has the same size as the y output.

to extrace from prameters given the specific HVIApproximation

Co-authored-by: Copilot <copilot@github.com>

Co-authored-by: Copilot <copilot@github.com>

so that can compute/check penalties  without providing observations.