Steepen backup-QB decay so the model stops drafting a second QB

Found via mock_draft: in a vanilla 1-QB, non-superflex 13-team league,
the model rostered a real backup QB (Mahomes/Nix-tier, not a late dart
throw) in 27 of 30 trials -- confirmed live after a user flagged the
same pattern showing up as "most common pick" across several rounds.

Root cause: QB draft_score is structurally much larger than every other
position's at mid-round draft slots (passing yards and TDs outscore the
field even for a QB nobody would roster twice), so the existing 0.20
backup multiplier, an 80 percent discount, still was not enough to push
a second QB's pick_value below real bench RB/WR value. Traced one actual
decision point: Patrick Mahomes (already had 1 QB) beat Stefon Diggs
6.52 to 4.54 despite the discount. 0.04 pushes every backup QB below
bench value in the same 30-trial check (0/30 drafted a second QB, vs
27/30 before).

TE's BACKUP_DECAY is untouched -- TE is flex-eligible in this league's
default settings (unlike QB), so a second TE can genuinely start via
FLEX, and the one 2nd-TE decision point checked was a near-tie against
real RB bench value (2.60 vs 2.55), not the lopsided QB case.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>