Discrepancy in Result Files for GPU-based Neuron Serialization

[d-kamath]

#569 CPU-based Neuron serialization operates as expected. We verified this by comparing two simulation approaches:

1. Running the entire configuration file in a single simulation (e.g., 10 epochs).
2. Running the same configuration in two stages:
   Stage 1: Simulating and serializing the first half of the configuration (e.g., 5 epochs).
   Stage 2: Deserializing the saved state from Stage 1 and completing the simulation with the second half (e.g., 5 epochs).
   The final result file from both the full simulation and Stage 1 & Stage 2 are identical, confirming the correctness of the serialization process.

However, for GPU-based simulations, while serialization is functional, the output result file from a full simulation and two half simulations differ. This discrepancy arises because the GPU’s random noise array is not serialized. To resolve this, we need to implement a method to manually copy the GPU state into temporary storage during serialization and restore it during deserialization.