Improvements to the moving weighted median

[jrs65]

• Support more than just double types. This is probably just as simple as switching to the Cython fused types
• Support n-dimensions. I can't decide if this one is hard or not.
• More options for dealing with missing data.
• Simplify logic by just inserting data with zero weights into the tree, and have the median return an appropriate value if all data is missing.
• Use enums for selecting the median method type. Chars are a bit ugly.
• Move the weight types into templates
• Split into a separate package. This would be pretty useful to the wider Python/Numpy community