Fix ANKH tokenizer to load from checkpoint hub id

[avivko]

FAST_ANKH_ENCODER always attached ElnaggarLab/ankh-base, so ANKH3 checkpoints (256-token vocab) were tokenized with the wrong token ids, generating NaNs in the embeddings. Load from config._name_or_path (same pattern as DPLM) with ankh-base fallback for bare configs.

Add slow tests that fast tokenizer ids match the official repo for ANKH_base, ANKH3_large, and ANKH3_xl.

Summary

• FAST_ANKH_ENCODER no longer hardcodes ElnaggarLab/ankh-base; it loads the tokenizer from config._name_or_path (e.g. Synthyra/ANKH3_large), matching the checkpoint vocab.
• Adds slow tests comparing fast vs official token ids for ANKH_base, ANKH3_large, and ANKH3_xl.

Hub follow-up

After merge, re-push modeling_ankh.py to Synthyra ANKH checkpoints (e.g. via get_weights.py) so trust_remote_code=True users pick up the fix without a local checkout.

Test plan

• pytest testing/test_ankh_tokenizer.py -v (Docker, 6 passed)

[avivko]

@lhallee An agent wrote this PR, sorry that it was a bit messy. But it fixes an important issue: I was getting NaNs in the embeddings when using ANKH3 and ANKH3XL. The tests for these models passed your original parity tests because you are using the native tokenizer for both in the parity test