Treat HRF convolution as a BIDS-Model transformation

[tyarkoni]

If HRF convolution is treated as just another BIDS-Model transformation, several non-trivial problems we currently face magically disappear (e.g., how to specify post-convolution transformations, how to apply different HRFs to different variables, etc.). This will require some work in the spec to introduce a controlled vocabulary for HRFs, at which point we can just add an HRF convolution transformation in pybids.analysis.transformations that wraps the nistats hemodynamic_models.compute_regressor function (see #250).