tflite: mtk_neuron: Respect the model_token provided by caller

The current model token generating method can easily generate thes same
token for different models. Since we're also providing the model token
and is expecting them to be unique, we can mix the original token
calculation logic and the provided token to avoid collision.

BUG=b:378633716
TEST=Run VC and not see cache corruption error

Change-Id: I0088c57d345769f45ae70977043362377e2f1920
Reviewed-on: https://chromium-review.googlesource.com/c/chromiumos/platform/tflite/+/6074030
Reviewed-by: Shik Chen <shik@chromium.org>
Tested-by: Tommy Chiang <ototot@google.com>
Commit-Queue: Tommy Chiang <ototot@google.com>

delegate/mtk_neuron/neuron_delegate_kernel.cc

diff --git a/delegate/mtk_neuron/neuron_delegate_kernel.cc b/delegate/mtk_neuron/neuron_delegate_kernel.cc
index 4d9d670..edcbecd 100644
--- a/delegate/mtk_neuron/neuron_delegate_kernel.cc
+++ b/delegate/mtk_neuron/neuron_delegate_kernel.cc
@@ -2175,11 +2175,12 @@
   }
   // gen model token
   nn_compilation_cache_token_.clear();
-  std::string bytes_count_str =
+  std::string model_token_str =
+      options_.model_token + "/" +
       std::to_string(TfLiteOpaqueContextGetNumTensors(context));
-  const char* model_token = bytes_count_str.c_str();
   uint64_t token_parts[4];
-  token_parts[0] = ::util::Fingerprint64(model_token, std::strlen(model_token));
+  token_parts[0] =
+      ::util::Fingerprint64(model_token_str.c_str(), model_token_str.size());
   if (params != nullptr) {
     token_parts[1] = GetHash(params->nodes_to_replace);
     token_parts[2] = GetHash(params->input_tensors);