(#2583) remove low_cpu_mem_usage from qwen init, it does not do anything anymore; use dtype instead of torch_dtype

[bghira]

This pull request improves the loading process for text encoder models in simpletuner/helpers/models/flux2/model.py. The changes optimize device placement during model initialization and add informative logging to help debug device allocation in distributed setups.

Device placement improvements:

• Changed model loading for both Qwen3 and Mistral-3 text encoders to load models onto CPU first, then move them to the per-rank accelerator device, preventing GPU contention during initialization. (_load_text_encoder_qwen3, _load_text_encoder_mistral) [1] [2]
• Updated model loading parameters by replacing torch_dtype with dtype and removing low_cpu_mem_usage for consistency and clarity. (_load_text_encoder_qwen3, _load_text_encoder_mistral) [1] [2]

Logging enhancements:

• Added detailed logging statements when moving Qwen3 and Mistral-3 text encoders to their target devices, including device information and process rank, to aid in debugging distributed training setups. (_load_text_encoder_qwen3, _load_text_encoder_mistral) [1] [2]