[Perf] Improve FP8 dispatch storage release

[A-nnonymous]

Summary

• enable fused y1 quant handoff in the FP8 gated path
• decouple down-projection backward from BF16 w2 storage in FP8 aligned mode
• update ernie compat FP8 weight attributes and call signatures to match functional APIs

Validation

• python -m py_compile sonicmoe/functional/__init__.py sonicmoe/ernie_compat/mlp_node_v2.py
• PaddleFleet import smoke: paddlefleet_ops.is_sonic_moe_available() == True, bare sonicmoe evicted from sys.modules
• 20-step PaddleFormers SonicMoE FP8 smoke completed on 8 GPUs with no Traceback/OOM
• 300-step run before final cleanup completed with stable current memory (~16.986 GiB) and global peak 79.953 GiB vs BF16 85.731 GiB

🤖 Generated with Claude Code

[A-nnonymous]

Added a minimal correctness fix in _DownProjection: restore ctx._topk_scores_needs_grad = True so Paddle torch-proxy does not silently drop router/topk score gradients.

Validation on Yichen-aligned 200-step configs (stopped after step75/80 as requested):

• BF16 Sonic step75 loss: 4.87523794 vs Yichen BF16 baseline 4.91020107
• FP8 Sonic offload step75 loss: 4.8631115
• Perf sanity: BF16 step75 runtime 3.4606s; FP8 step75 runtime 3.5747s with max_memory_allocated 111.0GB, so no material perf regression observed.

This intentionally keeps the diff to one source file / one logic change.