Add anagram word puzzle environment

[Thomasyoung113]

Adds a new environment for training LLMs on anagram/word unscrambling tasks. This tests pattern recognition and vocabulary skills.

Features

• Built-in word list with 100+ common English words of varying lengths (3-10+ characters)
• Configurable word length range via min_word_length and max_word_length
• Support for thinking mode with <think></think> tags
• Automatic word scrambling with guarantee that scrambled version differs from original
• Clean answer extraction with <answer></answer> tags
• Comprehensive metrics tracking (success rate, format compliance, etc.)
• WandB integration for visualization

Example

Given scrambled letters elppa, the model should respond with:

<think>Let me rearrange these letters... e, l, p, p, a... that's apple!</think>
<answer>apple</answer>

Configuration

Uses similar patterns to text_reversal_environment.py for consistency.