BFwf with many grads and laps + Hyperduals

[CheukHinHoJerry]

This branch is a super-branch of 3d-2 and 3d. Sorry for messing it up.

This branch is ready for review and it implements a basic BFwf_chain with:

• grad w.r.t X and parameter (with Zygote pb and Hyperdual)
• laplacian w.r.t. X (with Hyperdual)
• gradp laplacian (with Hyperdual + Zygote on gradp)

This should be enough to proceed after some performance fix. @cortner Please review this when you are free.

CC @DexuanZhou

[cortner]

I have an idea - I'll draft something, we can talk on Zulip.

[cortner]

See ACEsuit/Polynomials4ML.jl#45

[CheukHinHoJerry]

I removed the CI of J1.0 due to Julia version constrain by [email protected]. I also removed the ACEcore dependency, replace all of them by new P4ML and make sure all the tests are added back to run_test.jl (so that CI is actually meaningful) and everything works fine for now (at least from what I can tell).

TODO:

1. Remove DenseLayer implementation and use the one in P4ML
2. benchmarks again after things being rewrote in P4ML and swapping out the DenseLayer
3. there are some import conflicts that I should look into.

[CheukHinHoJerry]

Close and reference to #22