Address potential errors when using Poniard in environments with existing versions of dependencies

[rxavier]

• Makes the BaseShuffleSplit import more explicit, since it was not always a public class.
• Set minimum versions for Pandas, scikit-learn and XGBoost. These work with Python 3.7, so Poniard can still be used in Colab and Kaggle.
• Minor updates to setup.py.
• Streamlines dev requirements by keeping only ipykernel instead of the full Jupyter dep tree.
• Closes Prevent version errors with old preinstalled dependencies #3 #