[metalearn] implement raw metafeature encoder

[cosmicBboy]

An extension to the encoder that generalizes the metadata feature
representation from hand-crafted features (e.g. mean of means of numerical
features) and instead formulate encoder as a ANN function approximator module, where the input is some raw, numericalized form of the data, and the output is some lower-dimensional encoding of the data.

Possible architectures for the differentiable module:

Transformer

1. A hierarchical transformer that encodes each feature vector as "feature embeddings" and and each row vector as "instance embeddings".
2. feature embeddings are computed on random samples of the feature set {f_0, f_1, ... f_n}
3. instance embeddings are computed on random samples of the dataset (X, y)
4. the embeddings are concatenated and passed through a fully-connected layer that has the same output dimensionality as the controller input dimensionality.
5. this metafeature embedding is then used as input to the controller (decoder) to produce a ML framework.

CNN

A convnet can probably by used from the using the same data setup as the Transformer.