perf(krea2): use cuDNN SDPA with native GQA

[DualChimerra]

Summary

• replace Krea2's manual 12-to-48 K/V head expansion with native PyTorch GQA
• select the official cuDNN SDPA backend on CUDA when the current operands are supported
• retain normal PyTorch SDPA selection as a safe fallback for CPU and unsupported shapes
• document the effective Krea2 attention path

Validation

• 147 Krea2 and family integration tests passed
• 87 related modeling, sampling, and block-swap tests passed
• CPU native-GQA output matches explicit K/V expansion
• RTX 4070 Ti CUDA smoke test selected cuDNN SDPA for Krea2's 48/12 heads and completed forward/backward with finite outputs and gradients