Integer packing in fmpz_lll #2389
Description
When factoring S7 and S8, profiling shows about 40% of the time being spent in _fmpz_lll_check_babai on _fmpz_vec_scalar_submul_si calls updating the B matrix, whose entries are always one limbs up to a few limbs. It might make sense to switch to a homogeneous mpn representation for B here in some cases. Or is it possible to defer these vector updates and do them batched as more efficient matrix-matrix multiplies?