AINT_TR optimization

[talassio]

Subroutine AINT_TR

1. Deal separately with B generic and B=Identinity, I(n).
2. w%B is acting as a dense Identity matrix, and is mostly not required, except for the calls to solve_spd and solve_general, some proposals follow:
3. Assembly and allocation of m0(2n,2n) and m1(2n,2n) might not be required, specially m1 where it is constituted of two diagonals identities.
4. Change max_eig so not to expect m1 and m0 but the assembly blocks a and (potentially a non identity block b)
5. Solving ( A + lambda* B ) x = b can be optimized for the case where B=I(n) and probably the allocation of b(n,n) will not be required.

RALFit/libRALFit/src/ral_nlls_internal.f90

Lines 1560 to 1580 in df24780

	SUBROUTINE AINT_tr(J,A,f,X,v,hf,n,m,Delta,d,normd,options,inform,w)
	! -----------------------------------------
	! AINT_tr
	! Solve the trust-region subproblem using
	! the method of ADACHI, IWATA, NAKATSUKASA and TAKEDA
	! -----------------------------------------
	Implicit None
	REAL(wp), intent(in) :: J(:), A(:,:), hf(:), f(:), v(:), X(:), Delta
	integer, intent(in) :: n, m
	real(wp), intent(out) :: d(:)
	real(wp), intent(out) :: normd ! ||d||_D
	TYPE( nlls_options ), INTENT( IN ) :: options
	TYPE( nlls_inform ), INTENT( INOUT ) :: inform
	type( AINT_tr_work ) :: w
	integer :: i, size_hard(2)
	real(wp) :: obj_p0, obj_p1, obj_p0_gn, obj_p1_gn
	REAL(wp) :: norm_p0, tau, lam, eta
	Character(Len=80) :: rec(1)
	If ( .not. w%allocated ) Then