fix: fall back on older CUDA GPUs

[robinspt]

Summary

• disable FlashAttention automatically on pre-Ampere CUDA GPUs
• fall back from bfloat16 to float16 when bf16 is unsupported
• apply the compatibility fix to both the HF runner and end-to-end script

Why

On older NVIDIA GPUs, the current HF path can fail with:

• FlashAttention only supports Ampere GPUs or newer
• bf16 runtime errors on devices without bfloat16 support

This change keeps the existing fast path on supported GPUs while allowing older CUDA GPUs to run through the SDPA + fp16 fallback path.