perf(rl): wire fused reward pipeline — 7 launches → 1 + delete dead kernels
Replace extract_realized_pnl_delta + unit_state_update +
step_counter_update + abs_copy + reward_shaping + raw_rewards snapshot
+ recent_outcome_update with single rl_fused_reward_pipeline launch.
Delete 4 dead kernel handles (extract_realized_pnl_delta,
rl_recent_outcome_update, rl_reward_shaping, rl_step_counter_update)
from struct, constructor, and cubin consts.
b=128 benchmark: 5000 steps in 79.4s = 8,064 transitions/sec (4.4× baseline).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
f5d8f5f056