fix: Uses cuBLAS instead of the CUDA kernel to perform matmul

[henrique-borba]

• Creates NDArray::save($a, string $filename) and NDArray::load(string $filename) methods for persisting NDArrays as an alternative to serialize. This may be faster in some cases.
• Remove the matmul custom kernel and use cuBLAS. There does not appear to be a significant overhead that would justify the use of a specific kernel. Implemented by https://github.com/SkibidiProduction

[henrique-borba]

Fix #51