iRow and JCol values should be checked after TNLP::eval_jac_g call

[mottelet]

When TNLP::eval_jac_g is called with values==NULL (constraints Jacobian pattern is expected) there is no verification that the indices returned in iRow and jCol are compatible with n and m. For example if the sparsity pattern is recovered from another language (e.g. Scilab), indices start from 1 and if you forget to substract 1 from all values, then Ipopt crashes and the cause is very hard to debug (even with valgrind or ASAN).