This adds support for loading FLUX.2-klein models in HuggingFace diffusers format — users can load models directly from ~/.cache/huggingface/hub/ without converting to GGUF or BFL first.

Also fixes #699 — the split-fuse qkv fix is generic and resolves diffusers format loading for SD3, FLUX.1 and Z-Image as well.

Along the way I fixed two issues that affect all diffusers-format models, not just FLUX.2:

Split-fuse qkv loading. The .weight.N suffix convention in name_conversion.cpp was already used for SD3, FLUX.1 and Z-Image to split q/k/v into separate tensors, but load_tensors() never handled the split parts — they were silently dropped as "unknown tensor" and the base tensor had a shape mismatch. This PR adds dst_offset support so split parts get written into the correct positions of the fused destination tensor. This fixes diffusers loading for all models that use split qkv, not just FLUX.2.

Multi-shard safetensors. Large HF models are often sharded into multiple .safetensors files with a model.safetensors.index.json index. Added init_from_safetensors_index() that reads the index and loads each shard. This enables loading sharded text encoders (e.g. FLUX.2-klein-4B Qwen3 TE is 2 shards, FLUX.1-dev T5 is 2 shards).

Other changes:

• init_from_diffusers_file() now tries transformer/ before unet/ for DiT models
• Reads model_index.json to auto-detect LLM text encoders (Qwen, Llama) and use the correct prefix
• FLUX.2 diffusers adaLN modulation stores [shift, scale] vs BFL [scale, shift] — halves are swapped at load time

Before / After

Without this PR, loading FLUX.2-klein-4B diffusers format fails:

# Sharded text encoder:
WARN  file text_encoder/model.safetensors not found
ERROR load tensors from model loader failed

# Diffusers transformer:
ERROR get sd version from file failed

With this PR, both work:

# Sharded text encoder:
INFO  loaded 2 shards from 'text_encoder/model.safetensors.index.json'

# Diffusers transformer:
INFO  Version: Flux.2 klein
INFO  generate_image completed in 159.85s
INFO  save result image (success)

Testing

Tested on Apple Silicon (M3 Max, 36 GB):

• FLUX.2-klein-4B diffusers: all 418 tensors loaded, image generated, pixel-identical to BFL at same seed
• FLUX.2-klein-4B directory loading (--diffusion-model <dir>/): auto-detected Qwen3 TE, generated OK
• Multi-shard TE: FLUX.2-klein-4B (2 shards) and FLUX.1-dev T5 (2 shards) loaded
• Comparison at 256x256, 512x512, 1024x1024: diffusers vs BFL identical
• Regression: FLUX.1-dev GGUF Q8_0 (780 tensors, full generation OK), FLUX.2-klein-9B GGUF Q8_0 (201 tensors OK)
• 41 unit tests for name conversion (all passed)
• No new warnings in changed files

BFL (reference) vs Diffusers (this PR) — FLUX.2-klein-4B, 1024x1024, seed 42, 4 steps, prompt "gandalf you shall not pass":