[FEATURE] Save dataset_train.reader.class_to_idx inside the model and load it by default during inference and validation

**Is your feature request related to a problem? Please describe.**
I'm not using imagenet, but during inference it loads the imagenet class_map by default.

**Describe the solution you'd like**
Instead, the class_to_idx from the dataset_train.reader.class_to_idx should just be saved somewhere inside the model, and be loaded into class_map during inference by default. and of course, if someone still wants to override the class_map for whatever reason, they could still do so.

I'd make a PR myself, but y'all probably have other consideration for exactly where to save/load it, so here's a sample solution.

Sample solution:

```
# load
if args.class_map == ''
ckpt = torch.load(args.checkpoint, map_location='cpu')
args.class_map = ckpt.get('class_to_idx', '')

# save
output_dir = Path(output_dir)
best_model = list(output_dir.rglob("*best*"))[0]
ckpt = torch.load(best_model, map_location='cpu')
ckpt["class_to_idx"] = class_to_idx
torch.save(ckpt, best_model)

```
**Describe alternatives you've considered**
We could save the class_to_idx into a class map file and ship it along side the model, but that's cumbersome and tedious. The proposed solution just works by default.

**Additional context**
The same should probably be done with the args.yaml file. There are a ton of timm models on hugging face with pretrained weights, but no args.yaml file with them which makes it near impossible to reproduce their results.