Reduced M number for larger lambda, optional for both density and density-response

[zekunlou]

Since density fitting has less orbitals on large lambda auxiliary basis, reduced M number is reasonable, accelerating the construction and inversion of the Hessian matrix.
It is now only implemented for density-response and is not optional.
The plan is to enable optional reduced M number for both density and density-response, together with a complete test on model training cost and prediction performance.

Note:

• The key variable is Mcutsize. Print it when it is constructed.
• Introduce parameter inp.gpr.Mtype = "uniform" | "Gaussian".