Add ability to serialize Distributions and RegressionDatasets.

[akleeman]

This gives you the ability to serialize a RegressionDataset provided the FeatureType is also serializable. This can come in handy when it's costly to parse data and convert it into the required features and targets. For example, you can now you can do something like:

// Parse data and cache it.
std::vector<RegressionDataset<MyFeature>> datasets = parse_lots_of_data();
{
  std::ofstream ofs(cache_path);
  cereal::BinaryOutputArchive archive(ofs);
  archive(datasets);
}

Allowing you to read them in much faster next time,

// Read the cached datasets.
std::vector<RegressionDataset<MyFeature>> datasets;
{
  std::ifstream ifs(cache_path);
  cereal::BinaryInputArchive archive(ofs);
  archive(datasets);
}

[kmdade]

this hurts me :( I feel like we should extend the character limit on our formatter.

[akleeman]

yeah kinda annoying. I guess it's already a pretty difficult line to read, so doesn't make it that much worse :)

[kmdade]

you implicitly get SerializableDiagonalMatrix(const BaseClass &&other) I think, so that's the only scenario where you don't actually need to copy.

[kmdade]

lgtm.