[doc] feat: add effectiveness figures and example recipes to README

[Luosuu]

What does this PR do?

Concise overview of the change. Reference related issues/PRs.

Checklist Before Starting

• Search for relative PRs/issues and link here: ...
• PR title follows [{modules}] {type}: {description} format (see check_pr_title.py for the full list of allowed modules and types)
  • Breaking changes: prepend [BREAKING] — e.g. [BREAKING][ops] feat: new batch-invariant matmul API

Test

Validation results (numeric checks, benchmark metrics) for changes not covered by CI.

API and Usage Example

Show API changes and usage examples if applicable.

Design & Code Changes

High-level design description and specific change list.

Checklist Before Submitting

• Read the Contribute Guide
• Applied pre-commit checks (pre-commit run --all-files)
• Added/updated documentation
• Added tests to CI workflow (or explained why not feasible)

[gemini-code-assist]

Code Review

This pull request enhances the README.md by adding an 'Effectiveness' section with performance comparisons and an 'Example Recipes' section containing a comprehensive table of training scripts. The review feedback identified several areas for improvement, including ensuring model naming consistency, adding a required environment variable to an example command, and correcting an algorithm label in the recipe table to align with the actual script implementation.

[gemini-code-assist]

For consistency with the recipe table below (line 53), the model name should include the -Base suffix when referring to the REINFORCE++ experiment.

Suggested change

	> **Qwen3-30B-A3B · REINFORCE++ · DAPO dataset**
	> **Qwen3-30B-A3B-Base · REINFORCE++ · DAPO dataset**

[gemini-code-assist]

The example command is missing the test_path environment variable. The script examples/moe/run_qwen3_30B_A3B_dapo.sh explicitly requires test_path to locate the validation dataset (see line 36 of that script). Without it, the command will fail for users who do not have the default Arnold-style mount.

Suggested change

	model_dir=/path/to/model data_dir=/path/to/data bash examples/moe/run_qwen3_30B_A3B_dapo.sh
	model_dir=/path/to/model data_dir=/path/to/data test_path=/path/to/test bash examples/moe/run_qwen3_30B_A3B_dapo.sh

[gemini-code-assist]

The algorithm for this recipe is listed as DAPO, but the corresponding script examples/moe/run_qwen3_30B_A3B_dapo.sh sets algorithm.adv_estimator=grpo (line 34). Please ensure the algorithm name in the table accurately reflects the implementation in the script.

Suggested change

	| [`moe/run_qwen3_30B_A3B_dapo.sh`](examples/moe/run_qwen3_30B_A3B_dapo.sh) | Qwen3-30B-A3B | DAPO-Math-17k / AIME 2025 | 1×8H100 | DAPO |
	| [`moe/run_qwen3_30B_A3B_dapo.sh`](examples/moe/run_qwen3_30B_A3B_dapo.sh) | Qwen3-30B-A3B | DAPO-Math-17k / AIME 2025 | 1×8H100 | GRPO |