Allow user to pick which Text Encoders to train with SD3-Medium.

**Is your feature request related to a problem? Please describe.**
Kinda yes, it seems both training scripts for LoRA and Dreambooth load all 3 Text Encoders and Tokenizers which leads to high vram usage, 20GB on Rank 32 Lora, batch size 1, with AdamW8Bit

**Describe the solution you'd like.**
Allow users to train the Text Encoder(s) they want.

**Additional context.**

Start by locating the LoRA and Dreambooth training scripts for SD3-Medium and trace where all three Text Encoders and Tokenizers are loaded. Review the related training configuration and memory behavior first. Done means users can select which Text Encoder(s) to train while unselected encoders are not unnecessarily loaded, with coverage for the supported choices.

Scritto dal modello di indicizzazione a partire dal testo della issue.