Fix SFT for VLM example

[qgallouedec]

Closes #1786

For some reason, when running sft with llava 1.5 without PEFT, you get this annoying errors:

• loss explosion because of high grad norm in the visual encoder; you need to specify a low max_grad_norm (0.1 at least)
• when training with multiple GPUs, you get OOM error (even with batch size = 1). Not sure where it comes from

Consequently,

• I removed the example script without peft
• fixed the command args, in particular the use of bf16

TODO:

• Add a section in SFT documentation with link to run, command, a few remarks, and evaluations results

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[qgallouedec]

Not related to this PR:
CI breaks with transformers-4.43.1
(passes with 4.39.2)

[qgallouedec]

I'm still not sure how to evaluate, given that open-compass/VLMEvalKit is not compatible with llava 1.5. I'll do this later.
And the previous code for evaluation is not working:

## Evaluation:

To evaluate, first install the lmms-eval framework: pip install git+https://github.com/EvolvingLMMs-Lab/lmms-eval.git
then run:
accelerate launch --num_processes=8 -m lmms_eval \
    --model llava_hf \
    --model_args pretrained llava-hf/llava-1.5-7b-hf \
    --tasks mmbench \
    --batch_size 1 \
    --output_path ./logs/ \
    --log_sample

No module named 'llava.data'

Removing it until we find a working point