[Feature] Use 1D weight tensors for norm layers to match HuggingFace conventions

[lyfne123]

Summary

Change norm weight (gamma/beta) tensor parameters from 2D shape [1, H] to 1D shape [H] in Qwen3 model examples, to match the actual weight layout in HuggingFace model checkpoints.

Motivation / Use Case

HuggingFace model weights for RMSNorm and LayerNorm are stored as 1D tensors of shape [hidden_size]. Currently, Qwen3 examples define these weights as 2D tensors with shape [1, hidden_size]. This mismatch means users must reshape weights when loading from HuggingFace checkpoints, adding unnecessary friction for model deployment.

Affected files:

• examples/models/qwen3/qwen3_14b_decode.py
• examples/models/qwen3/qwen3_14b_prefill.py
• examples/models/qwen3/qwen3_32b_decode.py
• examples/models/qwen3/qwen3_32b_prefill.py
• examples/models/qwen3/qwen3_32b_training_draft.py

Proposed API / Behavior

Change norm weight parameter declarations from:

gamma: pl.Tensor[[1, HIDDEN], pl.FP32]

to:

gamma: pl.Tensor[[HIDDEN], pl.FP32]

This applies to all rms_norm_weight, input_norm_weight, post_norm_weight parameters in the Qwen3 model examples.