Change the default solver in sparse_generic_solve

[tvercaut]

We currently used minres as the default solver in sparse_generic_solve:

torchsparsegradutils/torchsparsegradutils/sparse_solve.py

Lines 407 to 416 in efa515c

	# ---------- default solvers ----------
	if solve is None and transpose_solve is None:
	from .utils import minres
	solve = minres
	transpose_solve = minres
	elif solve is None:
	solve = transpose_solve
	elif transpose_solve is None:
	transpose_solve = solve

While this makes sense for symmetric matrices and is documented, it may be confusing for new users who may expect sparse_generic_solve to mostly work out of the box even for non-symmetric matrices. Switching to BiCGSTAB would lead to a more intuitive behaviour.

Users who know their matrix is symmetric can always make the choice to specify a more specific solver such as minres. If we want to simplify that path, we can also create another thin wrapper along the lines of:

def sparse_generic_symmetric_solve(
    A: torch.Tensor,
    B: torch.Tensor,
    solve: Optional[Callable[..., torch.Tensor]] = None,
    **kwargs,
) -> torch.Tensor:
    # ---------- default solvers ----------
    if solve is None
        from .utils import minres
        solve = minres

    return sparse_generic_solve(A, B, solve=solve, transpose_solve=solve, **kwargs)