Add one-head nanoGPT Attention Residuals baseline

[charlesmartin14]

What changed

• Adds Full Attention Residuals (residual_mode: full_attnres) to the existing one-head nanoGPT model.
• Implements content-dependent depth routing with a zero-initialized learned pseudo-query and parameter-free RMS-normalized keys before both attention and MLP sublayers.
• Preserves separate W_Q, W_K, W_V, W_O, W_MLP_IN, and W_MLP_OUT matrices for Muon and WeightWatcher.
• Keeps AttnRes pseudo-query vectors in the auxiliary AdamW optimizer group rather than the 2-D Muon matrix group.
• Adds an exact 10-epoch matched AttnRes/Muon config using the same FineWeb-Edu data, context length, width, one block, one head, optimizer schedule, probes, and WW settings as the existing long-Muon baseline.
• Adds a conservative full-10-epoch cosine Muon schedule as a matched standard-vs-AttnRes pair. This is an experimental long-horizon candidate, not claimed as an empirically optimized schedule.
• Adds unit tests for forward/backward finiteness, routing weights, matrix separation, optimizer partitioning, and multi-depth readiness.
• Adds ATTENTION_RESIDUALS.md with controlled run commands and convergence-comparison criteria.

Why

The goal is to test whether learned residual routing accelerates convergence under exactly matched token/data conditions, while retaining the RG/WeightWatcher matrix instrumentation used by the existing nanoGPT baselines. Keeping an exact matched schedule and a separate matched long-cosine pair avoids conflating an architectural gain with a learning-rate-schedule gain.

Validation

The branch diff was checked for scope and the AttnRes equations were cross-checked against the Kimi Team Full-AttnRes formulation. Connector-only repository access in this session did not provide a local executable checkout, so the added pytest suite is included for CI/runtime validation rather than claiming a local test pass.