64873095

By: ValentinHirschi <valentin.hirschi@gmail.com>

masters: native B1, B11 and B111 for every carrier

The exact Passarino-Veltman B1, B11 and B111 divide by powers of p² and
lose 2-4 digits per decade of m²/|p²|: DoubleFloat and Float(128)
bubbles at small |p²| with unequal masses were silently wrong (relative
error up to 1.3e13 at rank 4). Route them through oneloopmaster's native
tensor families (B11, B111), with the per-row precision rule that B1111
already uses for Float. Dual carriers keep the exact forms.

V0 re-blessed: only 43 generic Float(256) components move, in their last
digit; fast and generic binary64 results are bitwise unchanged.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>