Issues related to new gradient formulation GRAD_FORM=2

[daniel-koehn]

There are still some issues related to the gradients for the non-conservative stress-velocity formulation (Castellanos, 2014) - GRAD_FORM = 2:

• The FWI only converges when using PCG optimization
• The inversion of hydrophone data is currently not supported, because the source excitation for the backpropagtion of the data residuals requires some modifications
• Compared to GRAD_FORM = 1 the ambiguity between Vs and density model is much higher. In case of the Marmousi-II model, structures from the Vs model dominate over density structures.

[daniel-koehn]

Solved the above mentioned issues by new gradients for stress-velocity formulation (GRAD_FORM=2) using a decomposed impedance matrix to introduce a self-adjoint forward operator. Notice, that no numerical integration of the data residuals is required anymore. While this improves the robustness of the FWI against noise, lower frequencies are required in the FWI workflow to mitigate trade-offs between different parameter classes. The data integration in the GRAD_FORM=1 formulation emphasizes lower frequency content in the data residuals. The formulations GRAD_FORM=3 and GRAD_FORM=4 are removed.