This daft PR introduces the base for iterative solvers.

Two methods are proposed:

• Conjugate Gradient (CG)
• Preconditioned Conjugate Gradient (PCG)

Each method is made public with two public interface flavors:

• solve_<method>_kernel: All arguments are mandatory (no optionals/no internal allocations), it contains the methods steps. The linear system (and preconditioner) is defined through a public DT linop which enables to extend two key procedures: apply equivalent to a matrix-vector product and inner equivalent to the dot_product. This is the interface recommended to extend the method when dealing with a matrix type not available in stdlib or when working in distributed-memory frameworks for which the matvec, dot_product and factorization steps need to be adapted to account for parallel synchronization.
• solve_<method>: API with optional arguments, the linear system can be defined with dense or CSR_<>_type matrices. For the PCG, the following preconditioners are available: none (identity), jacobi (1/diagonal). It internally uses solve_<method>_kernel.

• tests
• examples
• documentation