[TRAIN-8] Add end-to-end custom reward functions tutorial

[mkcash]

Summary

Closes #2724

Adds a new tutorial page Custom Reward Functions to the docs/guides section, providing an end-to-end example for:

What's included

1. Implementing reward functions — Two complete examples:

   • conciseness_reward() — scores responses by length
   • keyword_coverage_reward() — scores by required keyword presence

2. Registering via custom environment — ConcisenessEnvironment class that:

   • Subclasses BaseEnvironment
   • Uses the @register_environment decorator
   • Supports multi-reward (GDPO) and single-reward (GRPO) patterns
   • Accepts configurable parameters via env_config

3. Training config integration — Complete YAML example with:

   • trainer.env: "ConcisenessEnv" wiring
   • GDPO multi-reward weights
   • CLI override examples

4. Standalone runnable example — FormatAdherenceEnvironment as a minimal starting point

Related docs

• See docs/guides/environments.md for built-in environment reference
• See docs/guides/grpo.md for training configuration details

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.