Fix cuDNN LSTM implementation selection with LoadSavedModel C++ API.

[API92]

If I save tf.keras.layers.LSTM layer with _could_use_gpu_kernel=True into the SavedModel format with tf.saved_model.save, then cuDNN kernel is used when I load this model with tf.saved_model.load and it works fast. But if I load this model from C++ with tensorflow::LoadSavedModel function, then cuDNN kernel isn't used and it is slow.

Here is colab demonstrating the issue https://colab.research.google.com/drive/16WN0sqOoL37M7-5XMhGb-irkRX7fh503?usp=sharing . If model is loaded with tf.saved_model.load, then tf.keras.layers.LSTM and tf.raw_ops.CudnnLSTM layers both takes about 100 ms. But if model is loaded with LoadSavedModel from C++, then tf.keras.layers.LSTM takes about 275 ms, while tf.raw_ops.CudnnLSTM takes 100 ms.

There were some problems in FunctionOptimizer, ImplementationSelector and MetaOptimizer optimizers:

1. When FunctionOptimizer specialized some function, it copied api_implements from origin function to specialization. It caused an error "Function ... and ... both implment 'interface_name' but their signatures do not match" in FunctionLibraryApiInfo::Init
2. MetaOptimizer didn't pass whole FunctionDefLibrary along with a body of single function, which is needed for implementation selection.
3. ImplementationSelector couldn't parse empty device name if it isn't specified on a node.

[gbaned]

Hi @penpornk Can you please review this PR ? Thank you!

[gbaned]

Hi @ezhulenev Can you please review this PR ? Thank you!

[gbaned]

@API92 Can you please check build failures. Thank you!

[API92]

@gbaned Fixed.

[gbaned]

Hi @ezhulenev Can you please review this PR ? Thank you!

[penpornk]

Adding @reedwm since @ezhulenev is on vacation for a while.

[keerthanakadiri]

Hi @API92 , Can you please resolve the conflicts? Thank you!