Fix multi-threaded dataloader for Qwen3/Mistral text encoders

[BitcrushedHeart]

Description

Enables dataloader_threads > 1 for Z-Image and Flux2.Klein models by working around a thread-safety bug in the transformers library's check_model_inputs decorator (huggingface/transformers#42673).

Closes #1291

Problem

The check_model_inputs decorator in transformers v4 monkey-patches child module .forward() methods on every call to capture output_hidden_states, then restores them after. When two dataloader threads call the same text encoder concurrently, they race on patching/restoring these methods, causing hidden states from different threads to bleed into each other.

Fix

Wraps the text encoder's .forward() with a per-instance threading.Lock to serialize concurrent calls, preventing the race condition. The lock is applied conditionally only when dataloader_threads > 1 and is idempotent (safe if called multiple times).

Performance impact is negligible since GPU computation is already serialized on a single device. The benefit of multiple dataloader threads (pipelining CPU image loading/preprocessing against GPU encoding) is preserved.

Also proactively applies the same fix to the Flux2.Dev (Mistral) path, which has the same underlying vulnerability via MistralModel.forward().

The upstream fix (huggingface/transformers#43765) shipped in transformers v5 only. This workaround can be removed when upgrading to v5+.

Changes

• modules/util/thread_safety.py: New utility — apply_thread_safe_forward() wraps a model's forward with a per-instance lock
• modules/dataLoader/ZImageBaseDataLoader.py: Replace NotImplementedError with thread-safe forward patch
• modules/dataLoader/Flux2BaseDataLoader.py: Replace NotImplementedError (Klein) and add proactive fix (Dev)

Testing Notes

Verified the bug and fix with a tiny Qwen3ForCausalLM (CPU, random weights, 280K params):

• Without lock: 4 threads x 100 iterations — hidden states corrupt immediately (expected 3 layers, got 6-9)
• With lock: 4 threads x 100 iterations — 400/400 calls correct, zero errors

Tested on Windows 11, Python 3.10.11.

Will run a full test on Z-Image either today or tomorrow.

[BitcrushedHeart]

Tested on Z-Image. Using 12 threads took caching 110k files from 3.5 hours to 50 minutes.