[Misc] [ROCm] Prevent surplus tensor reshape (#19803)

Signed-off-by: Zsolt Borbely <zsolt.borbely@htecgroup.com>

Author: zsolt-borbely-htec

vllm/v1/attention/backends/triton_attn.py

@@ -376,7 +376,7 @@ def forward(
                     query.reshape(
                         (num_tokens, num_heads * head_size)).contiguous(),
                     layer._q_scale)
-            query = query.reshape((num_tokens, num_heads, head_size))
+                query = query.reshape((num_tokens, num_heads, head_size))
 
         use_local_attn = \
             (self.use_irope and attn_metadata.local_attn_metadata is not None)