fix: ignore submodule overrides when continuing fine-tune from backup

[dxqb]

fixes #1402

[dxqb]

Claude: clear_overrides() is too broad. It clears all six override fields, but two of them point at required components that aren't stored in the backup loader path, so resuming a fine-tune from backup will fail to load:

• Würstchen / Stable Cascade — the internal backup saves only the prior pipe (WuerstchenModelSaver.__save_diffusers → create_pipeline().prior_pipe). The effnet encoder and decoder come from effnet_encoder_model / decoder_model. After clearing them, the loader does DDPMWuerstchenScheduler.from_pretrained("") / load_file("") → raises. (The prior override is already neutralized by __load_internal, so it never needed clearing.)

• HiDream — the text_encoder_4 encoder weights fall back to the backup subfolder when empty, but tokenizer_4 loads unconditionally with no fallback (HiDreamModelLoader.py:113): LlamaTokenizerFast.from_pretrained("") → raises.

The safe fix is to clear only what the backup actually carries — transformer_model (and optionally vae_model, harmless). Leave effnet_encoder_model, decoder_model, and text_encoder_4 intact.