Revising the conda release strategy

[xiki-tempula]

I have a use case where I need some components of the pymbar that is not solver related like the decorrelation, so I install pymbar-core=4. In the same environment, I install openfe, which then pulls in pymbar=3. So in the same conda environment, I have pymbar-core=4 and pymbar=3, which is kind of bad.
I think it might be better to have jax and none-jax variant of pymbar as different build varient of pymbar?

Related to #500
I made a branch to show the solution that I came up with conda-forge/pymbar-feedstock#43

[mikemhenry]

I think the cleaner fix is to prohibit pymbar-core from being installed with pymbar=3.x (which we can do with a new build of pymbar 3 + a metadata repo patch).

I will let others chime in that might have some ideas @mattwthompson @Lnaden

[mattwthompson]

I don't really know what this package does, but in general if JAX can be treated as an optional dependency for some set of functionality, splitting the package out is a good idea