[WIP] Introduced pm.TidyData class

[kc611]

Link to Issue: #3953

As mentioned in the issue, added a pm.TidyData class for automatically translating strings from a xarray DataSet into integer arrays for indexing in models. The input (currently) is expected to be a xarray.Dataset.

This is just a basic implementation to see if it's going in the right direction. New feature suggestions are welcome. The expected usage is as follows :

import pandas as pd
import pymc3 as pm
import numpy as np
import arviz as az
import theano 
from pymc3.data import TidyData
df = pd.DataFrame({
    'treatment':['Carboplatin']*3 + ['Adriamycin']*4 + ['Fluorouracil']*3,
    'age':np.random.randint(40,80,size=10),
    'location':['Berlin']*6 + ['London']*4,
    'outcome':np.random.randn(10),
})
df.index.name = 'patient'

xdf = df.to_xarray()

with pm.Model() as model:
    data = TidyData(xdf,import_dims=['treatment','location','patient','outcome'])

    intercept = pm.Flat('intercept')
    treatment = pm.Normal('b_treat',sd=0.2,mu=0,dims='treatment')
    location = pm.Normal('b_location',sd=0.2,mu=0,dims='location')
    age = pm.Normal('b_age',sd=0.2,mu=0)
    mu = (intercept
        + treatment[data.get_indexed('treatment')]
        + location[data.get_indexed('location')]
        + age*data['age']
        )
    pm.Normal('y',mu=mu,sd=1,dims='patient',observed=data['outcome'])
    trace = pm.sample()
    tr = az.from_pymc3(trace=trace,model=model)

[kc611]

@AlexAndorra Have a look.