fix: disable ORT memory pattern to prevent RAM growth with varying sequences

ONNX Runtime's enable_mem_pattern retains peak-sized activation buffers
across inferences with different sequence lengths, causing steady RAM
growth in long-running processes.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: n24q02m

qwen3_embed/common/onnx_model.py

@@ -108,6 +108,9 @@ def _load_onnx_model(
 
         so = ort.SessionOptions()  # type: ignore[possibly-missing-attribute]
         so.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL  # type: ignore[possibly-missing-attribute]
+        # Disable memory pattern optimization to prevent ORT from retaining
+        # peak-sized buffers across inferences with varying sequence lengths.
+        so.enable_mem_pattern = False
 
         if threads is not None:
             so.intra_op_num_threads = threads