Add neural networks and bug fixes

[MikhailKuz]

Additions

• Neural networks in tabular preset (feature pipelines, models, tuning utils, grad clipping for regularization)
• Tutorial for nn and link for Kaggle competition
• Loss on logits for linear model (fixes Input contains NaN error when doing linear_l2 model #75)

Fixes

• kfold and reader seeds - if the sample (100k) is less than the number of observations per fold, then it samples the observations from only one fold (since folds are sorted). So the multiclass (and multioutput) target encoder with division by zero falls.
• deprecated series append - verbose deprecation warning (fixes Use of deprecated pandas methods #56)
• verbose - get right verbose level from verbose in automl.fit_predict
• weight in metric func - make available use of custom weight instead of default 1 in neural networks
• fix cb custom multitask metric - passed one-dim version of losses instead of multi-dim version (fixes Bug with custom metrics in multioutput tasks.  #74)

test

[dev-rinchin]

I'd like to place this section before section with TabularAutoML init

[dev-rinchin]

remove redundant variables DATASET_NAME and DATASET_DIR -- they are used just only once

[dev-rinchin]

Redundant comments in lines 105-107

[dev-rinchin]

please, add links to description of NN architecture for every section below

[dev-rinchin]

add definition for "swa"

[dev-rinchin]

Suggested change

	"#### 6.1.2 Define by yourself whole pipeline"
	"#### 6.1.2 Define the pipeline by yourself"

[DESimakov]

Remove all lambda functions.

Check whether you can load automl with nn on disk and load it.