Add custom chat template support for instructing tuning

[supreme-gg-gg]

Oops was doing checklist and realised we missed this rather important feature 💀

This is what is required to fine tuned pretrained base models... Will ship this ASAP!!!

from trl import SFTConfig, SFTTrainer
from datasets import load_dataset

trainer = SFTTrainer(
    model="Qwen/Qwen3-0.6B-Base",
    args=SFTConfig(
        output_dir="Qwen3-0.6B-Instruct",
        chat_template_path="HuggingFaceTB/SmolLM3-3B",
    ),
    train_dataset=load_dataset("trl-lib/Capybara", split="train"),
)
trainer.train()

[supreme-gg-gg]

Tasks:

1. Update the schema to accept a parameter called chat template path, if not provided we default to the gemma chat template (used for their instruction tuning by default -- well you should just use a IT model if that's the case lol)
2. If provided AND is base model, load the chat template by providing it in the SFTConfig (only SFT works with chat template obviously)
3. Update frontend to support this

We do not accept custom uploaded chat template (well you can always upload it to HF hub first, I would appreciate the courage of writing a chat template yourself)