Missing autorelease in lstm_mps caused a ton of leaked memory (pytorch#145503)

The dictionary held onto the new MPSGraphTensorData objects and MPSNDArrays.  Regression caused by pytorch#95137

Fixes pytorch#145374

Pull Request resolved: pytorch#145503
Approved by: https://github.com/Skylion007, https://github.com/malfet

Author: jhavukainen

aten/src/ATen/native/mps/operations/RnnOps.mm

@@ -322,13 +322,13 @@
     Placeholder outputPlaceholder3 = Placeholder(cachedGraph->outputTensors_[3], zState);
     Placeholder outputPlaceholder4 = Placeholder(cachedGraph->outputTensors_[4], cellStateFwd);
 
-    NSMutableDictionary<MPSGraphTensor*, MPSGraphTensorData*>* results = [@{
+    NSMutableDictionary<MPSGraphTensor*, MPSGraphTensorData*>* results = [[@{
       outputPlaceholder0.getMPSGraphTensor() : outputPlaceholder0.getMPSGraphTensorData(),
       outputPlaceholder1.getMPSGraphTensor() : outputPlaceholder1.getMPSGraphTensorData(),
       outputPlaceholder2.getMPSGraphTensor() : outputPlaceholder2.getMPSGraphTensorData(),
       outputPlaceholder3.getMPSGraphTensor() : outputPlaceholder3.getMPSGraphTensorData(),
       outputPlaceholder4.getMPSGraphTensor() : outputPlaceholder4.getMPSGraphTensorData(),
-    } mutableCopy];
+    } mutableCopy] autorelease];
 
     if (num_layers > 1) {
       Placeholder outputPlaceholder5 = Placeholder(cachedGraph->outputTensors_[5], layerOutputs);