fix(flux2): restore Mistral processor loading

[primorLee]

Summary

• restore AutoProcessor loading for the Mistral3 text encoder used by FLUX.2-dev
• retain AutoTokenizer for the Qwen3 text encoder used by FLUX.2 Klein
• add CPU regression tests for both model-family paths

Problem

The original FLUX.2-dev integration loaded its tokenizer directory through AutoProcessor. When Klein/Qwen3 support was added, the shared loader was changed to AutoTokenizer unconditionally.

FLUX.2-dev passes structured multimodal-style message content to apply_chat_template:

{"content": [{"type": "text", "text": prompt}]}

The Mistral3 path requires its processor to normalize that structure before tokenization. Calling the bare tokenizer diverges from both reference implementations:

• Black Forest Labs FLUX.2
• Hugging Face Diffusers FLUX.2

This explains why FLUX.2-dev prompt handling regressed while the Qwen3/Klein path still required a tokenizer.

Fix

Select the loader from the text encoder actually present in the model pool:

• Mistral3 / FLUX.2-dev -> AutoProcessor
• Qwen3 / FLUX.2 Klein -> AutoTokenizer

If both are absent, the default FLUX.2-dev configuration continues to use AutoProcessor.

Validation

• baseline current main: Mistral3 processor-selection test fails; Qwen3 tokenizer test passes
• after this change: 2 tests pass
• python -B -m unittest discover -s tests -p test_flux2_tokenizer_selection.py -v
• Python compilation passes for both changed files
• Ruff passes for the new test and the new production path after excluding ten unrelated pre-existing findings in the legacy pipeline file
• git diff --check passes

I could not run the gated FLUX.2-dev checkpoint end to end, so this PR claims restoration of the official prompt-processing path rather than an image-level parity measurement.

Closes #1324

[Artiprocher]

@primorLee Thanks for your contribution! We will fix it.