fix: upload tokenizer/config files with hub checkpoints

[salch-cred]

Summary

When a run's init checkpoint comes from one HF repo but checkpoints are uploaded to a different repo, the checkpoint repo only ever received the safetensors files — it was missing added_tokens.json, config.json (no— that one was sent), merges.txt, special_tokens_map.json, tokenizer.json, tokenizer_config.json, vocab.json and friends, so it was not loadable as a standalone model repository.

Fixes #631.

There were two gaps, fixed in two commits:

1. The download path never fetched these files.
MODEL_EXTENSIONS in both hub.rs and gcs.rs was [".safetensors", ".json", ".py"], so merges.txt, tokenizer.model (sentencepiece), and chat_template.jinja never made it to local disk in the first place. The two duplicated constants are consolidated into one MODEL_FILE_EXTENSIONS constant in file_extensions.rs (alongside the existing DATA_FILE_EXTENSIONS), extended with .txt, .model, .jinja.

2. The checkpoint extra-files filter missed several file names.
The checkpoint_extra_files filter in init.rs (duplicated in the hub and GCS branches) only carried config.json, tokenizer.json, tokenizer_config.json, special_tokens_map.json, generation_config.json, and .py files. Both duplicates are extracted into one is_checkpoint_extra_file helper carrying the full list from the issue (added_tokens.json, merges.txt, vocab.json) plus tokenizer.model and chat_template.jinja.

Matching is now by exact file name (via file_name()) instead of the previous path-suffix ends_with, which incidentally avoids false positives like a file named mytokenizer.json matching the old tokenizer.json suffix test.

Behavior change

Model downloads now also fetch tokenizer artifacts (a few small text/binary files — merges.txt is typically ~456KB, tokenizer.model ~500KB for a 32k vocab) and re-upload them with every checkpoint. Every caller of download_model_repo_async/_sync and download_model_from_gcs_* benefits with no API change.

Notes for reviewers

• All call sites of the old constants are updated (hub.rs x2, gcs.rs x1); MODEL_FILE_EXTENSIONS is also re-exported from the crate root next to the existing re-exports.
• The remaining .py match is kept for custom modeling code repos.
• Local checkpoint dirs (model_is_local branch reading an unpacked local repo) flow through the same helper, so a local init repo now also contributes its tokenizer files.

[salch-cred]

Note: I just noticed #647 addresses the same issue (#631) with a similar approach — apologies for the overlap, I searched open PRs for \checkpoint/\ okenizer\ but it didn't surface.

Differences vs #647, for the maintainers' convenience:

• This PR also fixes the GCS download path (\gcs.rs\ had its own copy of \MODEL_EXTENSIONS\ still missing .txt), while Preserve Hugging Face metadata when checkpointing #647 only touches \hub.rs\
• Consolidates the duplicated constants into one shared \MODEL_FILE_EXTENSIONS\ in \ile_extensions.rs\ (exported next to \DATA_FILE_EXTENSIONS), instead of growing two independent lists
• **Also covers \ okenizer.model\ (sentencepiece) and \chat_template.jinja**, which neither the old filter nor Preserve Hugging Face metadata when checkpointing #647 includes
• Preserve Hugging Face metadata when checkpointing #647 adds unit tests for the filter, which this PR doesn't — happy to add equivalent tests here, or to rebase this on top of Preserve Hugging Face metadata when checkpointing #647 if that one is preferred; whichever direction is easier for you.