fix: filter weight decay for LayerNorm, biases, and special tokens

[Luodian]

Summary

• Implement proper weight decay filtering following MAE/DeiT/OpenCLIP best practices
• Exclude 1D parameters and special tokens from weight decay

Problem

The current training applies weight_decay=0.05 to ALL parameters, including:

• 1D parameters: LayerNorm weights, RMSNorm weights, all biases
• Special tokens: learned probe, pos_embed, cls_token, etc.

This is suboptimal because:

• LayerNorm/RMSNorm scale parameters control feature scaling; decaying them reduces model capacity
• Biases don't benefit from weight decay (no overfitting risk from bias terms)
• Learned tokens like probe (shape 1,1,C) shouldn't be shrunk toward zero

Solution

Add build_adamw_param_groups() function that:

1. Shape-based rule: Exclude param.ndim < 2 (catches all 1D params)
2. Name-based rule: Exclude .bias suffix
3. Token-based rule: Exclude probe, pos_embed, cls_token, mask_token, query_tokens, latents

Evidence from Major Repos

Repository	Pattern
MAE	p.ndim == 1 + no_weight_decay_list
DeiT	timm-style no_weight_decay()
OpenCLIP	Two-group AdamW, WD=0 for p.ndim < 2
SigLIP	WD only on kernel (weight matrices)

Files Changed

• training/train.py - Add helper function and modify optimizer setup

Edge Cases Handled

• probe in pooling head: shape (1, 1, C) with ndim=3, explicitly excluded by name
• Conv2d bias: patch embedding uses bias=False, no issue
• MultiheadAttention biases: 1D, caught by ndim < 2 rule