Add Qwen3-30A3B AttnRes research experiment

[Randomizez]

What

This PR adds a research experiment for Attention Residuals on Qwen3-30A3B:

• new exp: playground/research/attn_res/qwen3_30a3b_attn_res.py
• supports:
  • attn_res_mode = "none" | "full" | "block"
  • block-mode residual routing with configurable attn_res_num_blocks
• keeps the existing Qwen HF reshaper path by subclassing QwenModel
• uses the existing Qwen3-30A3B SFT setup as the experiment base

Implementation notes

• AttnRes is applied at sub-layer granularity
  • one routing op before attention
  • one routing op before FFN
• effective AttnRes depth is therefore 2 * num_layers
• current implementation is restricted to:
  • PP=1
  • VPP=1
• for TP sequence parallel, the new pseudo_query parameter is marked so its gradients are synchronized across the TP group

Docs

The new experiment file starts with a short module doc that summarizes:

• the paper-style routing equations
• Full AttnRes pseudocode
• Block AttnRes pseudocode

Validation

Local checks performed:

• import/config smoke check
• cfshow config inspection
• scratch-init multi-GPU smoke runs using repo .venv/bin/torchrun

Scratch smoke runs after init cleanup were finite:

• seq_len=256: finite loss / grad norm
• seq_len=1024: finite loss / grad norm

Known limitations

• no real base checkpoint path is wired yet in this environment for Qwen3-30B-A3B-Base
• current PP support is intentionally disabled because AttnRes source-history semantics across pipeline stages are not implemented yet

Extra

This PR also records a few repo priors in AGENTS.md that came up while implementing and validating this experiment.