[src/MaxText/utils/ckpt_conversion] Split into one file per ckpt_conversion following strategy pattern

[SamuelMarks]

Description

1632 may just be too big a file! - See src/MaxText/utils/ckpt_conversion/utils/param_mapping.py

There are two obvious ways of doing this refactor:

0. Qwen, Gemma, Llama etc. files under big nouns like "checkpoint_conversion" or "multimodal_utils"
1. Qwen, Gemma, Llama with the same folder/module layout each, so Qwen3/checkpoint_conversion.py,
   Qwen3/get_image_offsets.py, Qwen3/apply_embedding.py, &etc.

If you take the latter approach to its extreme, the models themselves can be removed from this repository. New models can be added privately (e.g., on NVIDIA's or AMD's fork), and the loose coupling enables them to take full advantage of new/improved functionality throughout MaxText.

Various mechanisms can be made to make the second approach type safe, like:

• abstract classes https://docs.python.org/3/library/abc.html
• protocols https://typing.python.org/en/latest/spec/protocol.html
• adhoc approaches (
  e.g., JSON schema + LibCST + Pydantic)

And the original 0,1 can be done in phases. With 0 done first and 1 done once the final refactor can be done trivially performed in any IDE or langserv.

Even keeping all the models in this repository has advantage, particularly when refactoring / adding new features / upgrading dependencies and ensuring consistency throughout (for quality and benchmarking purposes).

Tests

CI

Checklist

Before submitting this PR, please make sure (put X in square brackets):

• I have performed a self-review of my code. For an optional AI review, add the gemini-review label.
• I have necessary comments in my code, particularly in hard-to-understand areas.
• I have run end-to-end tests tests and provided workload links above if applicable.
• I have made or will make corresponding changes to the doc if needed, including adding new documentation pages to the relevant Table of Contents (toctree directive) as explained in our documentation.

[codecov]

Codecov Report

❌ Patch coverage is 14.73684% with 486 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
.../MaxText/utils/ckpt_conversion/strategies/gemma.py	7.79%	142 Missing ⚠️
.../MaxText/utils/ckpt_conversion/strategies/llama.py	9.30%	78 Missing ⚠️
...axText/utils/ckpt_conversion/strategies/mixtral.py	8.00%	69 Missing ⚠️
...c/MaxText/utils/ckpt_conversion/strategies/qwen.py	13.15%	66 Missing ⚠️
...axText/utils/ckpt_conversion/strategies/gpt_oss.py	10.71%	50 Missing ⚠️
...xText/utils/ckpt_conversion/strategies/deepseek.py	17.94%	32 Missing ⚠️
...xText/utils/ckpt_conversion/utils/param_mapping.py	0.00%	30 Missing ⚠️
src/MaxText/integration/tunix/utils.py	0.00%	7 Missing ⚠️
...rc/MaxText/utils/ckpt_conversion/to_huggingface.py	0.00%	7 Missing ⚠️
src/MaxText/utils/ckpt_conversion/to_maxtext.py	0.00%	4 Missing ⚠️
... and 1 more

📢 Thoughts on this report? Let us know!