Set version to 2.0.1 (#376)

Backport: Fix activation lookup with Python 3.12.3 (#375) (#377)

* Fix activation lookup with Python 3.12.3 (#375)

We used the metaclass `EnumMeta`/`EnumType` to override reporting of
missing enum values (to give the full set of supported activations).
However, in Python 3.12.3, the default value of the `name` parameter of
`EnumType.__call__` method was changed from `None` to `_not_given`:

python/cpython@d771729

Even though this is a public API (which now uses a private default
value), it seems too risky to continue using it. So in this change, we
implement `Enum.__mising__` instead for the improved error reporting.

* Set version to 1.3.2

* Adjust two cross-tests for changes in HF transformers (#367)

* Fix `test_rotary_embeddings_against_hf` for latest transformers

* xfail test because HfFileSystem is currently broken

Finalize the API changes for 2.0 (#374)

* `qkv_split` arguments for attention heads are now mandatory

* Rename `FromHFHub` mixins to `FromHF`

* Remove `FromHF.convert_hf_state_dict`

Set version to 2.0.0.dev3, update curated-tokenizers dep to 2.0.0.dev0 (

#372)

* Update curated-tokenizers dependency to 2.0.0.dev0

* Set version to 2.0.0.dev3

* Set minimum version of Python to 3.9

Set version to 2.0.0.dev2 (#371)

Use small ELECTRA model for testing (#369)

Set version to 1.3.1 (#365)

AutoModel: let models check if the configuration is supported (#352)

* AutoModel: let models check if the configuration is supported

This will allow us to split up more complex models like Falcon into
multiple classes and corresponding entry points.

* Doc fix

Co-authored-by: Madeesh Kannan <[email protected]>

---------

Co-authored-by: Madeesh Kannan <[email protected]>

Update RTD links (#339)

Rename `download_to_cache` to `from_hf_hub_to_cache` (#323)