Add GSM8K examples for RLOO, REINFORCE++, GSPO and SAPO

[KunyangZhang]

Description

Four supported algorithms have no example config: RLOO, REINFORCE++, GSPO and SAPO
(their Doc / Example cell in docs/sphinx_doc/source/main.md reads -). This follows the
"Improve Examples" item in #470 ("Add examples for workflows or algorithms implemented but not
yet documented in the examples directory").

Background. The supported-algorithm table lists these four with algorithm_type values but no
example, so the only way to find a starting config for them is to read trinity/algorithm/algorithm.py.
Every other comparable algorithm (GRPO, CISPO, TOPR, sPPO, AsymRE, REC, OPD, JSD, ...) ships a
GSM8K example.

Changes.

• examples/rloo_gsm8k/, examples/reinforceplusplus_gsm8k/, examples/gspo_gsm8k/,
  examples/sapo_gsm8k/ — each with a gsm8k.yaml and a short README.md.
• The support tables in README.md, README_zh.md, docs/sphinx_doc/source/main.md and
  source_zh/main.md — fills in the example links for these four algorithms, and also restores the
  missing CISPO link (its example already exists in the repository but the table still showed -).
• docs/sphinx_doc/source/tutorial/example_dataset_perspective.md and the source_zh twin — adds
  the four GSM8K rows next to the existing CISPO/AsymRE rows.

Each YAML is examples/grpo_gsm8k/gsm8k.yaml with exactly two lines changed: the experiment name
and algorithm.algorithm_type. Dataset, model, cluster, buffer, explorer and trainer sections are
untouched, so the algorithm's own default_config() supplies the component wiring and the
resources stay at the smallest existing GSM8K footprint (2 GPUs).

How to test. The new configs are covered by the existing
tests/common/config_test.py::test_all_examples_are_valid, which walks examples/ and runs
load_config + check_and_update() on every YAML. I ran that exact body locally against the two
pre-existing GSM8K configs and the four new ones — all six pass identically. I also printed the
resolved wiring to confirm each example selects the intended components:

examples/rloo_gsm8k/gsm8k.yaml               algorithm_type=rloo                advantage_fn=rloo                 policy_loss_fn=ppo
examples/reinforceplusplus_gsm8k/gsm8k.yaml  algorithm_type=reinforceplusplus   advantage_fn=reinforceplusplus    policy_loss_fn=ppo   (gamma=1.0)
examples/gspo_gsm8k/gsm8k.yaml               algorithm_type=gspo                advantage_fn=grpo                 policy_loss_fn=gspo  (clip_range_low=3e-4, clip_range_high=4e-4)
examples/sapo_gsm8k/gsm8k.yaml               algorithm_type=sapo                advantage_fn=grpo                 policy_loss_fn=sapo  (tau_pos=1.0, tau_neg=1.05)

AlgorithmConfigValidator().validate(cfg) passes for all four, so the registry lookups and the
per-component default_args() resolution are exercised. I could not run pytest tests/common/config_test.py
directly here because the module imports vllm at collection time (vllm.config.ModelConfig) and this
machine has no GPU/vLLM install; the local run above reproduces the test body instead.

Checklist

• Code has passed all tests (example configs validated with the body of test_all_examples_are_valid)
• Docstrings have been added/updated in Google Style (no Python code changed)
• Documentation has been updated (algorithm support tables + dataset-perspective tables, EN and ZH)
• Code is ready for review

Refs #470

CI here runs on /unittest from a maintainer; I will watch for the result and fix anything it reports.