feat: Upgrade to latest TensorFlow Lite library (new GPU ops)

ios/TensorFlowLiteC.framework/Headers/builtin_ops.h

@@ -200,6 +200,23 @@ typedef enum {
   kTfLiteBuiltinStablehloBroadcastInDim = 170,
   kTfLiteBuiltinStablehloConvolution = 171,
   kTfLiteBuiltinStablehloSlice = 172,
+  kTfLiteBuiltinStablehloCustomCall = 173,
+  kTfLiteBuiltinStablehloReduce = 174,
+  kTfLiteBuiltinStablehloAbs = 175,
+  kTfLiteBuiltinStablehloAnd = 176,
+  kTfLiteBuiltinStablehloCosine = 177,
+  kTfLiteBuiltinStablehloExponential = 178,
+  kTfLiteBuiltinStablehloFloor = 179,
+  kTfLiteBuiltinStablehloLog = 180,
+  kTfLiteBuiltinStablehloMinimum = 181,
+  kTfLiteBuiltinStablehloNegate = 182,
+  kTfLiteBuiltinStablehloOr = 183,
+  kTfLiteBuiltinStablehloPower = 184,
+  kTfLiteBuiltinStablehloRemainder = 185,
+  kTfLiteBuiltinStablehloRsqrt = 186,
+  kTfLiteBuiltinStablehloSelect = 187,
+  kTfLiteBuiltinStablehloSubtract = 188,
+  kTfLiteBuiltinStablehloTanh = 189,
 } TfLiteBuiltinOperator;
 
 #ifdef __cplusplus

ios/TensorFlowLiteC.framework/Headers/xnnpack_delegate.h

@@ -35,6 +35,10 @@ extern "C" {
 // Enable XNNPACK acceleration for VAR_HANDLE, READ_VARIABLE, and
 // ASSIGN_VARIABLE operators.
 #define TFLITE_XNNPACK_DELEGATE_FLAG_VARIABLE_OPERATORS 0x00000010
+// Enable transient indirection buffer to reduce memory usage in selected
+// operators. Indirection buffer initialization will take place on every
+// inference run, instead of only once during initialization of the operators.
+#define TFLITE_XNNPACK_DELEGATE_FLAG_TRANSIENT_INDIRECTION_BUFFER 0x00000020
 
 struct TfLiteXNNPackDelegateWeightsCache;
 
@@ -48,6 +52,7 @@ typedef struct {
   // - TFLITE_XNNPACK_DELEGATE_FLAG_FORCE_FP16
   // - TFLITE_XNNPACK_DELEGATE_FLAG_DYNAMIC_FULLY_CONNECTED
   // - TFLITE_XNNPACK_DELEGATE_FLAG_VARIABLE_OPERATORS
+  // - TFLITE_XNNPACK_DELEGATE_FLAG_TRANSIENT_INDIRECTION_BUFFER
   uint32_t flags;
   // Cache for packed weights, can be shared between multiple instances of
   // delegates.