Allow lazily evaluated TruncateSVD results

[eric-czech]

I don't see any reason why the pca-specific parameters like these need to be computed immediately rather than leaving them as dask arrays:

dask-ml/dask_ml/decomposition/truncated_svd.py

Lines 181 to 187 in b94c587

	components, ev, evr, sv = compute(v, explained_var, explained_variance_ratio, s)
	self.components_ = components
	self.explained_variance_ = ev
	self.explained_variance_ratio_ = evr
	self.singular_values_ = sv
	self.n_features_in_ = X.shape[1]
	return X_transformed

Is that true @TomAugspurger? I'd like to add a compute flag if that's ok. I'm guessing there is nothing in scikit-learn that would ever rely on those specific attributes.