Instabilities and unexpected behavior with `FullInfiniteProjector`

[pbrehmer]

Here, I will document numerical instabilities and unexpected behavior with FullInfiniteProjector algorithm for future reference.

• In the fixed_iterscheme.jl test, FullInfiniteProjector leads to invariant subspaces that make the IterSVD fail for a (1, 1) unit cell. Here's the warning:

 Warning: Invariant subspace of dimension 1 (up to requested tolerance `tol = 1.0e-14`), which is smaller than the number of requested eigenvalues (i.e. `howmany == 16`).
└ @ KrylovKit ~/.julia/packages/KrylovKit/i0JjG/src/eigsolve/svdsolve.jl:182

• In the jacobian_real_linear.jl test, FullInfiniteProjector somehow isn't even linear for real coefficients. In particular, real_err_∂A and real_err_∂x are finite and will not drop below the error tolerance.