Replace src/scitypes.jl with ScientificTypes.jl package

[ablaom]

The first three changes below resolve MLJ issue #176

• (breaking for model implementations) Replace the model trait input_scitype_union with input_scitype. It's value is now some scitype defined by the new ScientificTypes package, eg Table(Continuous) (instead of Continuous) so that non-tabular data can be supported. See ScientificTypes documentation for more on the Table scientific type.

• (breaking for model implementations) Replace the model trait target_scitype_union with scitype_union with similar remarks applying.

• (breaking for model implementations) Eliminate target_is_multivariate and input_is_multivariate traits as this information is now implicit in value of above revamped traits

• (breaking) Eliminate scitypes method. The scientific types of a table are returned as part of ScientificTypes schema method, which is re-exported by MLJBase. (MLJBase.schema used to coincide with Tables.schema.)

• add package_license_trait (fallback value is "unknown")

• add supports_weights trait for measures and models

• Efficiency and other improvements to UnivariateFinite distribution (resolves Use LittleDicts in UnivariateFinite? #22), including implementation of Distributions.support. This is different from classes which returns all categorial values for the distribution, not just those with non-zero probability. This is breaking because classes used to return what support does now.

• Address DataFrames depreciation warning due to column access change in its API

• (breaking) CSV is now an optional dependency, which means you now need to import CSV before you can load tasks with load_boston(), load_iris(), load_crabs(), load_ames(), load_reduced_ames(). Addresses Implement the MLJ model API without needing to depend on external dependencies such as CSV.jl, CategoricalArrays.jl, etc. #19