Cache: use batch_size instead of max_batch_size

[gante]

What does this PR do?

Renames the input argument max_batch_size, present in static-shaped caches, to batch_size. max_batch_size is imprecise: the cache needs the EXACT batch size being used.

The imprecise variable name and description was a source of issues, e.g. here.

NOTE: while it is technically feasible to accept smaller batch sizes in static-shaped caches, we would be slicing the cache at each layer. Slicing is an expensive operation, and the point of using static-shaped caches is to be fast. In other words, we would be enabling a silent incorrect usage of the class 🤗

---

✅ all chances are backwards compatible, and the user will only see a warning if passing the batch size through the deprecated keyword argument

from transformers import AutoConfig, StaticCache

config = AutoConfig.from_pretrained("gpt2")

# No warnings
StaticCache(config, 8, 100, "cpu")
StaticCache(config=config, batch_size=8, max_cache_len=100, device="cpu")

# Warnings
StaticCache(config=config, max_batch_size=8, max_cache_len=100, device="cpu")

[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.

[zucchini-nlp]

Great, thanks for making it clearer!

[ArthurZucker]

thanks for updating, we could keep the name and make sure the description is better as well!

[ArthurZucker]

code on the hub will complain but yes it makes sense