Dataset Handling and coercion

[olivierlabayle]

Need to have a better management of the dataset and scientific types coercion.

In particular it needs to be explicit and parameterizable. Probably in TMLE.jl.

[ ] For coercion, because each column must be processed based on its role in the "causal graph" I can think of multiple levels:

1. Based on estimand dispatching on required nuisance functions
2. Based on nuisance functions
3. Make sure they agree if a variable is found in multiple such functions

[ ] For the dataset management.

1. A copy of the relevant coerced data will be forwarded through the estimation procedure
2. It is still unclear how the cache should be optimally handled by users (think security)

[olivierlabayle]

This has been partly handled now, now the main question is whether full dataset management should be ported to TMLE.jl. In the end it seems to be a simple coerce!(dataset, autotype(dataset)) typeof thin, potentially restricted to columns defined by an estimand.