[Perf] Enable SonicMoE FP8 storage release

[A-nnonymous]

Summary

• install native 1x32 SonicMoE FP8 weight caches from PaddleFleet expert quantization
• clear SonicMoE FP8 weight caches once per step before reinstalling to avoid stale tensor-version cache growth
• keep imports under paddlefleet_ops.sonicmoe and harden unavailable cudnn frontend loading
• update sonic-moe submodule to [Perf] Improve FP8 dispatch storage release PFCCLab/supersonic-moe#43

Validation

• python -m py_compile src/paddlefleet/transformer/moe/moe_expert.py packages/paddlefleet_ops/src/paddlefleet_ops/__init__.py
• import smoke: paddlefleet_ops.is_sonic_moe_available() == True, bare sonicmoe evicted from sys.modules
• FP8 config smoke: enabled=True, quack=True, fused_gated=True, fuse_y1_quant=True, fp8_wgrad=True, iso32_weight=False, recompute_z=False, save_z_fp8=False
• 20-step PaddleFormers SonicMoE FP8 smoke completed on 8 GPUs with no Traceback/OOM
• 300-step validation completed before final cleanup: FP8 global peak 79.953 GiB vs BF16 85.731 GiB; current memory stable around 16.986 GiB

Dependency

• Requires [Perf] Improve FP8 dispatch storage release PFCCLab/supersonic-moe#43

🤖 Generated with Claude Code