investigate we can run standard MCMC for non-galaxy image chains

[ismael-mendoza]

say for the toy interim posterior ellipticity and shear, could enable us to use CPUs which is much cheaper and for this low dimensionality faster?

• for interim ellipticity posterios --> OK , need to play with step size depending on noise ofc, but relatively fast using random walk mcmc from balckjax in cpu

• for shear posterior, probabiy OK, need to play with step size

• figure out a way to adapt step size i.e. cov_matrix

[ismael-mendoza]

it's better to keep NUTS with low number of max_doublings to avoid having to do the adaptation, even in CPU

note however that CPU performance serioulsy degrades if k is large