Fix svd

[zhucaoxiang]

Somehow, the present code cannot normally execute SVD regularization. I checked the source code and made two changes:

1. override Nlambda before current_potential related variables are declared;
2. re-allocate RHS matrix with a dimension of ntheta_plasma*nzeta_plasma, rather than num_basis_function.

@landreman Please check if I did something incorrectly.

[landreman]

Yeah, it looks like I broke the svd scan on my May 10 commit "Extracted prepare_solve, solve, diagnostics to eliminate repeated code," since I missed the fact that RHS has a different size for the svd_scan compared to the usual solves.

Can I ask why you're using the SVD scan? As shown in the regcoil Nuclear Fusion paper, the SVD method seems pretty bad compared to standard regcoil.

regcoil_build_matrices.f90 seems like a surprising place to set nlambda = num_basis_functions-1. Could we do that in, say, regcoil_compute_lambda.f90 instead?

[zhucaoxiang]

Can I ask why you're using the SVD scan?

It's just my personal interest. I was wondering if I can find a numerically stable solution with minimum regularization. I understand that the current density regularization has been proven to be much better than the others, especially when we care about the maximum currents flowing in the coils. However, I was thinking the SVD regularization might be more suitable for my purpose. We can just leave the singular values that are large enough, say, abandon those lambda_i < 10^{-3} * lambda_max. So I just tested it and found SVD is not working.

Could we do that in, say, regcoil_compute_lambda.f90 instead?

There is no particular reason for me to put nlambda = num_basis_functions-1 in regcoil_build_matrices.f90. I found it was in regcoil_build_matrices.f90 that num_basis_functions is defined. I think any place after num_basis_functions is defined and before current_potential related terms are allocated is fine. You can definitely put them in regcoil_compute_lambda.f90 for consistence.