WIP: Reduce memory footprint of key generation

[mkannwischer]

This PR works towards reducing memory consumption of key generation.

[mkannwischer]

I can't get the CBMC proofs to work for at this point.
I tracked down the problem and isolated it in mld_compute_t0_t1_foo.

With this workaround, I can pass the CBMC proofs:

    mld_poly ttt0;
    mld_poly ttt1;
    const mld_polyvecl *row = mld_polymat_get_row(mat, i);
    mld_compute_t0_t1(&ttt1, &ttt0, row, s1, &s2->vec[i], t);
    t0->vec[i] = ttt0;
    t1->vec[i] = ttt1;

If I change that to the actual code, it spins forever:

    const mld_polyvecl *row = mld_polymat_get_row(mat, i);
    mld_compute_t0_t1(&t1->vec[i], &t0->vec[i], row, s1, &s2->vec[i], t);

[mkannwischer]

I'm running out of ideas for this one.
@rod-chapman, do you happen to have some spare cycles to take a look at it, please?