[V1] Enhance check when clicing encoder output

Prepare for vllm-project#11409

For pixtral model, we need to insert placeholders in the middle of
encoder output, to fit into whole soft embedding. This case makes
slicing operation tricky. This PR raises assertion if something's
off.

Signed-off-by: Linkun Chen <[email protected]>

Author: lk-chen

vllm/v1/worker/gpu_model_runner.py

@@ -734,6 +734,7 @@ def _gather_encoder_outputs(
                 assert req_id in self.encoder_cache
                 assert i in self.encoder_cache[req_id]
                 encoder_output = self.encoder_cache[req_id][i]
+                assert end_idx <= encoder_output.shape[0], f"{end_idx=} {encoder_output.shape=}"
                 encoder_outputs.append(encoder_output[start_idx:end_idx])
         return encoder_outputs