[Bug]: TarsModels do redownload embeddings

[helpmefindaname]

Describe the bug

related but not the same as: #3167

Locally saving a TarsModel won't save the huggingface config and therefore requires internet connection & a longer loading time when loading it on a new machine without hf-cache.
This happens due to tars pickeling the internal SequenceTagger/TextClassifier instead of just seralizing their embeddings.

To Reproduce

from flair.models import TARSClassifier

model = TARSClassifier.load("tars-base")
model.tars_embeddings.model.config._name_or_path = "bert-base-uncased"
model.tars_embeddings.base_model_name = "bert-base-uncased"
model.tars_embeddings.name = "transformer-bert-base-uncased"
model.save("local-tars-base.pt")

# clear huggingface cache or copy `local-tars-base.pt` to another machine or docker container.

model.load("local-tars-base.pt")

Expected behavior

The model should load without the need of internet and it shouldn't require me to wait until the embedding config and weights are downloaded.

Logs and Stack traces

Downloading (…)okenizer_config.json: 100%|██████████| 28.0/28.0 [00:00<00:00, 17.2kB/s]
Downloading (…)lve/main/config.json: 100%|██████████| 570/570 [00:00<00:00, 517kB/s]
Downloading (…)solve/main/vocab.txt: 100%|██████████| 232k/232k [00:00<00:00, 2.12MB/s]
Downloading (…)/main/tokenizer.json: 100%|██████████| 466k/466k [00:00<00:00, 3.61MB/s]
Downloading pytorch_model.bin: 100%|██████████| 440M/440M [00:35<00:00, 12.5MB/s]

Screenshots

No response

Additional Context

No response

Environment

Versions:

Flair

0.12.2 (master branch)

Pytorch

2.0.0+cu117

Transformers

4.28.1

GPU

True