Remove _metadata_log in metadata implementation

This was always a hypothetical use case. The idea was that metadata would keep log of who and when changed it, so that at the end you could get a nice record of changes through the pipeline. But now I think it is not a good idea anymore to attach this to metadata itself:

• It requires all primitives to correctly link input metadata with output metadata for this to works, all in the pipeline.
• It makes metadata objects larger and larger as they accumulate those changes
• The unexpected side effect is that if you try to pickle metadata, it also goes to pickle source primitives, which are potentially fitted instances which have a lot to pickle. Which can even lead to errors: sklearn-wrap#91 (closed)
• It is probably making things slower then necessary.

I think it is much better if TA2 would collect those into a series of input and output metadata as it is running the pipeline. Then it can also exactly mark what was input and what output, and even record that two input metadata objects produced one output metadata object and so on. Or, if TA2 does not care about any of that, it does not have to do it and have efficient and sleek metadata.