Treating states properly, second encounter

[YichengDWu]

I just realized that the recent PR #604 did not fix anything. As Phi is a mutable field of immutable PhysicsInformedNN. The related source code is a bit messy. Adapting states should be done inside PhysicsInformedNN. But there are too many situations to deal with and I don't want to fix it right away. For me, I just want to assume that the input chain is always a Lux.Chain, not vectors. If there are multiple dependent variables, then wrap them with another Chain, like Chain(u = chain_u, v = chain_v). We can use keys(chain)==depvars to check if there are multiple outputs. The advantage of doing this is that we can initialize the parameters and states at once. Then discretization.Phi should probably be a named tuple so we don't need to remember the order of dependent variables.

[ChrisRackauckas]

I think that's reasonable. This was special cased before Chain could do that, and now that it can, we should aggressively simplify the code and interface.

[YichengDWu]

It was a bad idea to wrap them in a new Chain. I was not thinking correctly. It should've been just a named tuple because initialstates and initialparameters both work on named tuples, and we can nicely call phi.u.

[YichengDWu]

I guess simply putting Setsfield.@set! there can correctly mutate Phi.

[YichengDWu]

Moved to #687