Refactor HQC phase 4 — drop non-constant-time sparseMul

Restore dense Karatsuba multiplication at the four sites previously
using GF2x.sparseMul. The phase 1 sparseMul kernel gave a 4-6x speedup
but wrote to scratch buffer offsets derived from secret support
indices, regressing cache-line side-channel resistance versus the
previous dense path. Drop the kernel and revert the four call sites to
materialise the support through writeSupportToVector + gf2x.mul.

Retain the phase 2 ReedMuller cleanup and the phase 3 addSupportTo
fusion at the three addTo-only sites (x in genKeyPair, e and r1 in
pkeEncrypt); both preserve L1+L2+L3 constant-time properties and net
~3-6% end-to-end speedup over the original code across Java 8 through
Java 25.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>