Fix Gemma4 MoE Triton flex stage config

[FurtherAI]

Summary

• add the Gemma4 MoE compile-crash signature for 512-wide global attention
• carry Gemma4 flex compile crash config through the attention wrapper into copied layer configs
• make signature-selected Triton num_stages=2 override forced-Triton backend selection

Testing

• direct compile-flag probes passed
• scratch CP worker passed for tp1/ep2/cp2/no-sp; generated 512-head kernels used num_stages=2
• ruff check and ruff format passed
• commit hooks passed

No new tests are included.