Add Caching To The Estimators

Some kind of Feature Request. I propose to add caching approach to the library estimators ...

Let's consider how it can boost Feature Selection approach for example:

• This need/idea just stolen from heamy package :). - http://heamy.readthedocs.io
• Once you train estimator, it's cached automatically by default.
• In our case we can start from running SBS(k_features=(10,20)), evaluate
• AND next iteration SBS(k_features=(13,30)) will be performed much faster, because we will load our estimations (13,20) from the cache
• Heamy has one known for me cons - when I changed the dataset, but shape was not changed their caching algorithms can't recognize the difference :)

Thank you.

Start by reviewing the library's estimator implementations and the SBS feature-selection entry point mentioned in the issue. Clarify how trained estimators and dataset changes should be identified before defining the cache behavior; done means repeated SBS ranges reuse valid estimations without confusing changed datasets.