Optimize MaxText unit and integration test suite runtime

[shralex]

This pull request optimizes the MaxText TPU/GPU CI test suites to substantially reduce total execution duration. It eliminates redundant compilation, graph tracing, and setup latencies without affecting functional verification or reducing test coverage.

Classes of Optimizations:

1. Model Downscaling (Compilation): Scaled down model configurations (embedding dimensions, heads, layers) across unit and integration tests to minimize XLA compilation overhead.
2. Sequence & Step Capping (Execution): Capped sequence lengths at 128 or 512 (down from 1024/8192) and reduced loop steps/step counts across test assertions and runs.
3. Cached Data & Model Pipelines (Initialization & Compilation):
   Replaced method-level dataset recreation with class-level lazy caching to construct data pipelines once per suite.
4. Moved input pipeline tests to be cpu_only.
5. Added class-level caching of unquantized base model results in QuantTest to eliminate redundant compilation and forward/backward pass execution across multiple quantization tests.
6. Local Path & Tokenizer Redirection (I/O & Network):

• Enforced synthetic datasets and redirected GCS asset paths to local mock directories to eliminate network/GCS latency.
• Redirected remote Hugging Face tokenizers (google-t5/t5-large, deepseek-ai/DeepSeek-V3, and zephyr-7b-beta) in HF/Grain tests to local pre-saved asset directories (tokenizer.default and qwen3-tokenizer), completely removing external network request overhead.

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!