Supports general submodels and a bunch of new parameters

[Larsvanderlaan]

Changes:

1. Supports general submodels (see submodels.R and the .submodel private field in Param_base, and changes to tmle3_Updater). The most straightforward way is to specify an error distribution for the loss_function and a link function for the submodel. This can be done with R family objects.
2. Added very minor support for weights in the updater. It is grabbed from the tmle_task if present.
3. User can specify bounds for each node in Updater.
4. A made some changes to tmle3_Updater in how the universal submodel updating is performed. The collapsing of the clever covariate is moved to the fit_submodel function so the full-dimensional epsilon vector is stored/cached. Also EIF components are grabbed from clever covariates to do the collapsing if present/available. Ideally, clever_covariates() should return an item "EIF" which is a node-named list of EIF components. This is not necessary nor leads to any changes in the output for most parameters in tmle3.
5. A bunch of new conditional treatment effect parameters in both semi and nonparametric settings including:

• Param_spCATE: Semiparametric estimation of CATE assuming user-specified parametric model for CATE is correct
• Param_spOR: Semiparametric estimation of conditional odds ratio (OR) assuming user-specified parametric model for OR is correct
• Param_spRR: Semiparametric estimation of conditional relative-risk (RR) assuming user-specified parametric model for RR is correct. (e.g. relative risk regression)
• Param_npOR: A nonparametric working model version of spOR. Viewed as a best approximation of the true OR using the working model.
• Param_npCATE: A nonparametric working model version of spCATE where the true CATE is projected using least-squares onto a linear working model. If formula = ~1, this returns the nonparametric ATE estimate.
• Param_npCATT: A nonparametric working model version of spCATE where the true CATE is projected using least-squares onto a linear working model using only the observations with A = 1 (among the treated). If formula = ~1, this returns the nonparametric ATT estimate. This method is a more direct generalization of Param_spCATE but it should be interpreted as an approximation of the true CATE that predicts best among the treated.

[Larsvanderlaan]

Still some minor testing to be done on the new parameters. I have checked coverage of the methods with a sample formula = ~1+W model and parametric learners.