Flatten the means from BaeysianNeuralFieldEstimator.predict

[fsaad]

The predicted mean array returned by https://github.com/google/bayesnf/blob/2debb69cf21daf7de982843a82436a57d5a97dd3/src/bayesnf/spatiotemporal.py#L273

has three dimensions, (num_devices, ensemble_size // num_devices, len(table) (for MLE/MAP) or even four dimensions (for VI).

It is more convenient to have only two dimensions.

Related #19.