Packed basis matrices in LLL

[fredrik-johansson]

Allow fmpz_lll to use packed mpn representation of basis matrices instead of fmpz. Fixes #2389. For the multiprecision approximate LLL, few-limb dot products in the Gram basis are moreover done exactly via the dot product kernels introduced in #2824.

This turned out to have a bigger impact than I anticipated; factoring S10 got 1.7x faster. Improved timings on the hard polynomials in build/fmpz_poly_factor/profile/p-factor_hard -hard (the easier polynomials are unaffected):

                  old       new    speedup
H1              1.047     0.744     1.41x
S7              0.019     0.018     1.06x
S8              0.187     0.162     1.15x
C1              0.081     0.076     1.07x
S9              1.875     1.425     1.32x
H2              0.342     0.330     1.04x
P7*M12_5       27.568    21.049     1.31x
P7*M12_6       48.307    34.689     1.39x
S10            65.694    38.124     1.72x

Implemented using Claude Fable 5.1.

Also fixes a bug in fmpz_lll_heuristic_dot (apparently was harmless, but a bug nonetheless).