Add unit tests for Gemma 4 LoRA targeting and NNX caching

[RexBearIU]

Description

This follow-up PR adds robust, CPU-friendly unit tests for the Gemma 4 LoRA support introduced in #3969.

It implements the precise test coverage requested by reviewers to ensure:

1. LoRA Targeting Integrity: Verifies that the regex pattern defined in lora_module_path.yml for gemma4 correctly resolves and matches the expected module paths (including MoE paths).
2. MoE NNX Graph Caching Safety: Verifies that the NNX graph caching is dynamically disabled for MoE models (expert count > 1) to allow correct metadata synchronization, and remains enabled for standard Dense architectures (expert count <= 1).

Why this is a good solution:

These lightweight unit tests execute in milliseconds on standard CPUs, fully guarding these critical codepaths against future regressions without adding TPU dependency or CI execution overhead.

Related PR: #3969

Tests

To verify these changes, run the following pytest commands:

# 1. Run the SFT trainer model caching tests
pytest tests/post_training/unit/train_sft_test.py

# 2. Run the LoRA regex path matching tests
pytest tests/post_training/unit/lora_utils_test.py

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

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!