Fix: Support --train_text_encoder in train_dreambooth_sd3.py by handling None tokenizers/text_input_ids

[quchenyuan]

What does this PR do?

This PR fixes the --train_text_encoder flag in train_dreambooth_sd3.py, which previously failed during training because the encode_prompt and _encode_prompt_with_t5 functions did not properly handle cases where tokenizers or text_input_ids are None.

When --train_text_encoder is enabled, the training pipeline pre-tokenizes prompts and passes text_input_ids_list directly to avoid retokenizing in every step — but the original code assumed tokenizers was always available, causing crashes like:

AttributeError: 'NoneType' object has no attribute 'encode'
or
ValueError: text_input_ids must be provided...

This PR:

✅ Makes _encode_prompt_with_t5 and encode_prompt robust to None tokenizer by:

• Accepting precomputed text_input_ids as fallback
• Adding validation when tokenizer is missing
• Preserving batch size logic even when prompt is None

✅ Ensures compatibility between training and inference code paths

Fixes # ( 8507)