CUDA 3 support, drop OptimKit, add diagonal Hessian updates in MAP_joint

[marius311]

Aggregate of a bunch of things:

• Drop OptimKit and LBFGS entirely. The LBFGS just wasn't significantly better, sometimes worse actually. Additionally, https://github.com/Jutho/OptimKit.jl/issues/11 was just proving to be too annoying. Instead, MAP_joint now does some simple isotropized diagonal-only updates to the Hessian after a pre-determined burnin, which in practice I find works much better than LBFGS ever did.

• Adds a new function get_max_lensing_step that prevents trying to take a ϕ step so large that the subsequent magnification matrix is singular which would then cause LenseFlow to error.

• Rewrites MAP_marg to only depend on lnP, so if that's changed, its self-consistently refelected in the MAP_marg result .

• Adds CUDA 3 support, drops CUDA 2 support.

• Drops julia 1.5 support.

[marius311]

@kimmywu I'm ready to merge this but can you check it doesn't break any of your marginal MAP stuff first and approve it when done?

Fyi, on NERSC, I'm now running with module load cuda/11.3.0 and export JULIA_CUDA_MEMORY_POOL=binned

I double checked with https://github.com/marius311/marginal_MAP_example/blob/master/marginal_MAP_example.ipynb and it runs fine (and MAP_marg is almost ~2x faster). I also updated the MAP_marg code to only depend on lnP, so if thats changed the MAP_marg code will be self-consistent, as well as combining the old δlnP_δϕ into MAP_marg for simplicity.

[marius311]

Awesome, thank you for checking quickly. Will merge shortly.