fix: move quantized norm to CPU instead of stale q_linear reference in smooth_quant

[Mr-Neutr0n]

Summary

• Fix a copy-paste bug in lmdeploy/lite/apis/smooth_quant.py where the norm quantization loop calls q_linear.to('cpu') instead of q_norm.to('cpu')
• q_linear is a stale variable from the previous linear quantization loop, so the quantized QRMSNorm objects are never moved to CPU
• This causes a VRAM leak: every quantized norm layer stays on GPU, which can lead to OOM errors on large models with hundreds of norm layers

Details

In the smooth_quant function, there are two quantization loops:

1. Linear quantization loop -- correctly calls q_linear.to('cpu') after quantizing each linear layer
2. Norm quantization loop -- creates q_norm via QRMSNorm.from_float() but then calls q_linear.to('cpu') (copy-pasted from the linear loop) instead of q_norm.to('cpu')

The fix is a one-line change: q_linear.to('cpu') -> q_norm.to('cpu') in the norm quantization loop.

Test plan

• Verify smooth quantization completes without OOM on a model with many norm layers
• Confirm GPU memory is properly released after each norm quantization step