feat: enable dynamic cpu-offloading for qwen image based models

[Cyclica-bin]

Motivation

This PR automates CPU offloading in NunchakuQwenImageDiTLoader, replacing the manual block-count parameter with a dynamic offloading system.

Currently, users must manually specify the number of blocks to load into VRAM. This leads to:

• Trial and error: Users have to guess how many blocks fit their specific GPU.
• Crashes: If other processes (or browser tabs) consume VRAM, the static block count causes OOM (Out of Memory) errors.
• Speeds up the workflow execution even without --lowvram flag if user has sufficient RAM.

Modifications

1. Removal of num_blocks_on_gpu: Removed the manual input parameter from the ComfyUI node since users no longer need to guess how many transformer blocks their free VRAM can hold.
2. API Update on load_model: Removed the same parameter from the argument list of load_model method.
3. Uses Comfy UI's model management utilities to
   • Calculate the available free memory of GPU
   • The size of transformer per block calculation
   • Identifies the blocks that must be reserved in the memory that are variable size based on nature ( minor perf improvements )
4. Added additional clamping logic just in case there are architectural changes to base model revision to avoid unexpected crashes.

Checklist

• Code is formatted using Pre-Commit hooks (run pre-commit run --all-files).
• Tested on a hardware with 8GB VRAM with 32GB RAM.
• Tested on a hardware with more than 8GB VRAM.

Note: Other checklists are not applicable as this is an extension to existing node.

• For reviewers: If you're only helping merge the main branch and haven't contributed code to this PR, please remove yourself as a co-author when merging.
• Please feel free to join our Discord orWeChat to discuss your PR.