SVC massively overfits on multiclass data sets?

[rhiever]

In one of my latest commits, I removed the SVC model operator because I noticed it was massively overfitting on the MNIST data set. For example, it would achieve a 99% accuracy on the internal TPOT validation set, but would drop to ~30% accuracy on the external validation set. I'm not sure if this is a known issue with SVC, but I'm simply removing it in the meantime to prevent this from happening.

Hopefully the sklearn-benchmarks project that we're working on will shed some light on what's going on.