ENH: refactor Xdawn and linear_regression_raw

Xdawn and linear_regression_raw overlap a lot. Both use toeplitz to construct a big predictor matrix from event arrays, and return the coefficients of regressing the brain data. Possibly, linear_regression_raw could be extended slightly to function as a backend for Xdawn, as discussed with @alexandrebarachant .

The main issues:

• Xdawn has an option to return the predictors, although it's not used within the function I think.
• linear_regression_raw takes raw input, Xdawn epochs (although it seems the original intent was for Xdawn to take raw too) is probably the most complicated step.
• Allowing Xdawn to call linear_regression_raw, and to read the result, should be straight-forward.

I don't think it would make much sense to refactor in the other direction (use Xdawn as the estimator for linear_regression_raw, or fully lose linear_regression_raw in favour of big Xdawn API enlargement) due to API and usage case differences

Maybe this should wait until #2331 has been addressed though.

@alexandrebarachant

Start by comparing the Xdawn and linear_regression_raw implementations, especially their toeplitz predictor construction, input shapes, and returned coefficients. Review the discussion around #2331 before deciding whether linear_regression_raw can serve as Xdawn's backend; done means the overlap is reduced without breaking the distinct APIs or expected raw and epoch inputs.