Add support to re-estimate Smaps if trajectory changes during training

[chaithyagr]

We need a way to update the smaps internally while learning k-space trajectory.
This can get complicated and it is worth starting a discussion on this.

Basically, we need a way to estimate and update Smaps internally in between forward.op, which is calibrationless fourier operator and fourier_op.adj_op, which is a self-calibrated fourier operator..

[paquiteau]

Well, for sustainin the discussion, my questions are:

• Do we really want to do that ?
  From my understanding, the smaps have very low-frequency supports, and the center of k-space is not what's changing the most during trajectory learning (?)

• Also, if (when) we support autodiff for the smaps as well (I think that is of some interest for @AsmaTANABEN ) how would we organize this ?

• I didn't understand your last point , could you provide some pseudo code to clarifies things a bit ?

[chaithyagr]

• Yes, for non-Cartesian sampling in center of k-space. We need to re-estimate. The question is how do we do it right, esp as we need kspace_data to estimate smaps . Ideally we dont need this issue and we can expect users to actually have do it right themselves.
• We potentially wont need autodiff wrt Smaps anytime soon
• Well:

fourier_op.samples = new_samples
forward_op.samples = new_samples
Y = forward.op(multi_image) # to simulate the scanner, so we give multi channel image as input
update_smaps(fourier_op, Y)
recon = fourier.adj_op(Y)