fix: rewrite curiosity training to per-block reduce — eliminates non-deterministic atomicAdd
Replace warp-reduced atomicAdd gradient accumulation in the curiosity
forward model with a two-kernel deterministic pipeline:
1. curiosity_fwd_bwd_per_block: each block reduces its threads' gradient
contributions via shared memory tree reduction, writes one partial
gradient vector [CUR_TOTAL_PARAMS] per block.
2. curiosity_grad_reduce: one thread per parameter sums block partials
in fixed order (block 0, 1, 2, ...). Fully deterministic.
Deleted kernels: curiosity_forward_backward (atomicAdd path),
curiosity_fused_zero_fwd_bwd_adam (grid-wide atomic barrier),
curiosity_adam_step_fused (dead code), warp_sum_cur helper.
Rust side: removed block_counter, adam_fused_func,
fused_zero_fwd_bwd_adam_func. Added partial_grads buffer
[max_blocks * CUR_TOTAL_PARAMS] (~2.8 MB for 64 blocks).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
020e9ba460