Retrain EspalomaCharge with resonance-invariant features

[jchodera]

Thanks to @ljmartin in #16, we realize we used a resonance-sensitive feature atom.GetTotalValence() in the training of EspalomaCharge 0.0.8.

I've opened this issue so we can track progress toward the following updates:

• Retraining EspalomaCharge using the resonance-invariant atom.GetDegree() instead
• Removing the ability for users to specify net charge in charge()
• Adding tests to make sure the total partial charge matches the net formal charge (to within a tolerance)
• Adding tests to make sure carboxylates and other resonance-equivalent atoms end up with the same partial charge
• Releasing a new version