* merge changes for clipa inference

* update get_tokenizer to pass CI test; replace gelu_appoximate with act_kwargs

* Temporary, cannot have a force tf dependency

* Supporting SigLIP and CLIPA-v2 models (both sourced from big_vision jax based modelling code).

* Fix some test failures, remove old v1 CLIPA configs, add add 336 H14 CLIPA

* Fix torchscript

* Fix CoCa expand typo, force final LN after attentional pool

* Used wrong default clean fn in SimpleTokenizer, put lower case back

* Attempt to fix xlm roberta test w/ pretrained hf weight difference

* SigLIP weights working. More changes to support differing image preprocessing / text tokenization sensibly.

* A typo and unused import

* Fix two small issues, add hf_tokenizer_name to SigLIP models for non hf-hub use

* CLIPA reference temppory rwightman/ models for testing

* Rename profile->profiler to avoid python naming conflict

* More tokenizer rework, add context_len as class attr set in factory, default __call__() arg to None. Clean up reduction masking logic and fix #680

* fix ViT-SO400M-14-SigLIP name

* Fix CoCa pool LN, improve clarity of ViT pooling logic

* Exclude first/last tokens from tokens output of text models, should match prev CoCa behaviour, but at odds with argmax which leaves special tokens in (not consistent)

* Add eval results for CLIPA + SigLIP models

* Fixup bigG CLIPA config, 83.03 top-1 IN-1k

---------

Co-authored-by: zw <26880977+zw615@users.noreply.github.com>
Co-authored-by: Gabriel Ilharco <gabriel.ilharco@gmail.com>