Remove OpenNMT and Tensorflow, Update to Python 3.10

[TaperChipmunk32]

-Removed OpenNMT and Tensorflow dependency -Updated from Python 3.8 to Python 3.10 -Updated from Ubuntu 20.04 to Ubuntu 22.04 -Updated Transformers to >=4.38.0,<4.46

---

This change is 

[Enkidu93]

@Enkidu93 Will removing the WER scorer mess up Clowder or the onboarding app in any way?

Reviewed 29 of 29 files at r1, all commit messages. _Reviewable_ status: :shipit: complete! all files reviewed, all discussions resolved (waiting on @mshannon-sil)

No, it should be fine.

[TaperChipmunk32]

@TaperChipmunk32 Can you remind me why we can't upgrade to Transformers 4.46? From #521, it sounds like 4.46 supports SDPA for NLLB.

Reviewable status: all files reviewed (commit messages unreviewed), all discussions resolved (waiting on @mshannon-sil)

In this PR for transformers 4.46.0, they deprecated the parameter "tokenizer" for Trainer, renaming it to "processing_class". That parameter was used in several places in machine.py.

I'll look into its uses in silnlp, but updating to 4.46 should be very straightforward. It just may limit us to only >=4.46.0.

[isaac091]

It should be the default

[TaperChipmunk32]

@ddaspit The options for attn_implementation are eager, sdpa, or flash_attention_2. I am going with eager as the default, and a config option can be added later to change this.

[TaperChipmunk32]

Reviewed 1 of 1 files at r6, all commit messages. Reviewable status: all files reviewed, 1 unresolved discussion (waiting on @TaperChipmunk32)

_silnlp/nmt/hugging_face_config.py line 778 at r6 (raw file):_

            id2label={},
            num_labels=0,
            attn_implementation="eager",

Can you set attn_implementation on the config? In the documentation, they set it on PreTrainedModel.from_pretrained.

Thank you for catching this, I meant to put it in AutoModelForSeq2SeqLM.from_pretrained and not the config. It may be valid in AutoConfig, but I am not sure. I've moved it to match the documentation.