allow replacing batchnorm with groupnorm for pre-trained image encoder

[sean1295]

What this does

This PR simply allows replacing batch norm layers with group norm layers for pre-trained image encoder (like imagenet pre-trained weights) for diffusion policy, which is a standard practice, by treating it as a warning instead of raising an error.

How it was tested

Tested with lerobot.scripts.train.py. See below

How to checkout & try? (for the reviewer)

HF_ENDPOINT=https://hf-mirror.com/ CUDA_VISIBLE_DEVICES="0" python -m src.lerobot.scripts.train
--policy.type=diffusion
--policy.pretrained_backbone_weights="ResNet18_Weights.IMAGENET1K_V1"
--policy.crop_shape="[224, 224]"
--policy.num_train_timesteps=16
--policy.num_inference_steps=16
--dataset.repo_id=sean1295/pusht_256_filtered
--dataset.use_imagenet_stats=True
--batch_size=128
--steps=200000
--wandb.enable=true
--save_freq 5000
--log_freq 100
--job_name=pusht_dp_imn
--policy.push_to_hub=False