Fix for CB incosistency for qwen2_5_vl (#765)

For Qwen2_5_vl the `decode_inputs["position_ids"][decode_batch_id]` is
of size (4,1) and the code was only updating the pos_ids of last index
of last array. Therefore, changing it to update the last idx of all
arrays of all the batches.

---------

Signed-off-by: asmigosw <asmigosw@qti.qualcomm.com>

Author: asmigosw

QEfficient/generation/text_generation_inference.py

@@ -956,7 +956,7 @@ def run_continuous_batching_decode(self, prompt_queue, generation_len):
                 else:
                     # If the generated sequence is valid and within generation len prepare for next decode
                     decode_inputs["input_ids"][decode_batch_id, -1] = next_token_id[decode_batch_id, -1]
-                    decode_inputs["position_ids"][decode_batch_id, -1] += 1
+                    decode_inputs["position_ids"][decode_batch_id][..., -1] += 1
                     self.generated_ids[batch_id_map[decode_batch_id], generated_id_current_index[decode_batch_id]] = (
                         next_token_id[decode_batch_id, -1]
                     )