Fix FLUX2 Klein load-time VRAM spikes on low-memory GPUs.

[mhofer1976]

Keep the transformer and Qwen text encoder off CUDA during initial load/quantization in low-VRAM mode so model startup avoids full-model OOM before offloading and quantization can take effect.

Summary

• avoid moving the full FLUX2 transformer to CUDA before quantization, which caused startup OOM on low-memory GPUs
• keep the transformer/text-encoder on CPU during low-VRAM model preparation and only move as needed
• use qtype_te for Klein Qwen text-encoder quantization instead of the transformer qtype

Test plan

• Reproduce prior failure on FLUX2 Klein 9B at Loading transformer with low VRAM settings
• Verify loader no longer calls full-model CUDA move before quantization in the FLUX2 path
• Verify Klein TE path no longer eagerly loads full TE to CUDA and uses qtype_te
• Run a full training smoke test on a low-VRAM GPU and confirm model loads and begins training

[inflamously]

Omg yes please, anything that fixes these annoying VRAM spikes between sampling and at random steps. I have a 5090 RTX that can run this flux 2 klein 9b without any problems but the longer the run the more it spikes.

[jaretburkett]

Thank you @mhofer1976