I decided to redo the TSA like the RPA, ie like a pipeline of transformers.
I created three new classes of transformers for transfer learning Tl in tangent space Ts :
TlTsCenter, TlTsNormalize and TlTsRotate.

I used prefix TlTs, to make the difference between
new transformations in tangent space: TlTsCenter, TlTsNormalize and TlTsRotate;
and already existing transformations in manifold : TLCenter, TLStretch and TLRotate.

To be coherent, we could change the prefix of transformations in manifold,
using TlMa to emphasize that transfer learning Tl is applied in the manifold Ma.
TLCenter would be deprecated and renamed into TlMaCenter,
TLStretch into TlMaStretch, and TLRotate into TlMaRotate.

Another solution, maybe better, is to have a single class TlRotate which makes the rotation in the manifold if the dimension of input is 3 (like TLRotate), and in the tangent space if dimension is 2 (like TlTsRotate). Same for TlCenter, and TlScale (grouping TLStretch and TlTsNormalize).

What do you think @agramfort , @sylvchev and @plcrodrigues ?

Documentation of TLDummy, TLEstimator, TLClassifier and TLRegressor have been updated to be compatible in both TL cases, in manifold (as before) and in tangent space.
They are deprecated and respectively renamed into TlDummy, TlEstimator, TlClassifier and TlRegressor

To be coherent, TSclassifier is deprecated and renamed into TsClassifier.

Thanks @plcrodrigues for your feedback. It's a good idea not to duplicate classes.

In the lastest version, there is a single class TlCenter which makes the centering in the manifold if the dimension of input is 3 (like TLCenter), or in the tangent space if dimension is 2 (like TlTsCenter).
Same for TlScale, merging TLStretch and TlTsNormalize; and for TlRotate, merging TLRotate and TlTsRotate.

I complete documentation.

Ok for me to keep prefix TL, instead of Tl.