Describe the Issue

This issue addresses two critical defects in the model initialization logic within example_trainer/model.py:

1. Meta Tensor Traversal Bug: The get_parent_and_name helper was incorrectly using getattr for numeric parts of a tensor name (e.g., layers.0.weight). Since ModuleLists do not support string-named attributes for indices, this caused model initialization to crash for every transformer architecture.
2. RoPE Theta Desync: The framework hardcoded the RoPE base frequency to 10,000. Modern models like Llama 3 (500k) and Qwen (1M+) require significantly higher base frequencies. Using the wrong theta causes a complete breakdown of positional understanding at long contexts.

Environment/API Details

• Environment Class/Name: example_trainer/model.py
• Environment Configuration: Any configuration using Llama 3, Qwen, or other high-context models.
• API Endpoint/Method Involved: _initialize_meta_tensors and get_parent_and_name

Steps to Reproduce

1. Initialize a model with a ModuleList structure (e.g., Llama).
2. Observe AttributeError in get_parent_and_name when it tries to access numeric indices.
3. If manually bypassed, observe "stuttering" or gibberish output on contexts > 2048 due to theta mismatch.

Interaction Details (if applicable)

• Input Item to collect_trajectory:

  # Example of a broken path that fails:
  full_name = "model.layers.0.self_attn.q_proj.weight"
  # getattr(model.layers, "0") fails

• Expected Behavior:
  1. get_parent_and_name should detect numeric indices and use __getitem__.
  2. RoPE theta should be dynamically detected from the model's config.json.

Setup Details

• OS: Linux
• Python Version: 3.10+
• Atropos Version: commit c20c852
• Relevant Libraries/Versions: torch, transformers

Additional Context & Logs

This fix implements a multi-source RoPE theta detection logic that scans rope_theta, rope_base, and rope_scaling dictionaries in the model configuration.