Fused Add + RMSNorm pattern

[AndreSlavescu]

Motivation:

Qwen3 Decoder Layer has the following pattern in the forward pass:

hidden_states = residual + hidden_states

# Fully Connected
residual = hidden_states
hidden_states = self.post_attention_layernorm(hidden_states)

https://github.com/huggingface/transformers/blob/0dc2df5ddafe3cb5824ad24e85beba13e0aa6726/src/transformers/models/qwen3/modeling_qwen3.py#L271

Goal:

Design a fused cutlass kernel that performs the above pattern.

[AndreSlavescu]

PR for above:
#55