feat: add configurable Qwen cache types and early cache scheduling

[leejet]

Summary

• Add qwen_image_2_1_prefix_cache_type with automatic, floating-point, and supported quantized storage types using the existing type parser.
• Automatically use FP16 only when every attention operation selects Flash Attention; rebuild the prefix in FP32 if attention falls back.
• Centralize early cache-output scheduling in GGMLRunner and migrate Qwen, SenseNova, ControlNet, and Wan/LTX VAE callers to avoid retaining large source activations until graph end.

Related Issue / Discussion

Fix #2040.

Additional Information

N/A

Checklist

• I have read and confirmed this PR follows the contribution guidelines.