Enable GPU-supported RGF calculation by adding 'rgf_device'

[AsymmetryChou]

This pull request introduces comprehensive improvements to device management and tensor placement throughout the NEGF codebase. The main focus is to allow explicit control over which device (CPU or CUDA) is used for the recursive Green's function (RGF) step, while ensuring all relevant tensors and operations are consistently placed on the correct device. This is achieved by propagating an rgf_device parameter and updating tensor allocations, conversions, and file I/O accordingly. Additionally, the pull request clarifies device usage in documentation and argument parsing.

The most important changes are:

Device management and propagation:

• Added an rgf_device parameter (defaulting to CPU) to core classes and functions, including DeviceProperty, NEGF, and argument parsing, to control where the RGF step runs. All relevant tensor allocations and computations in the NEGF pipeline now use this device. [1] [2] [3] [4] [5] [6]

Consistent tensor placement:

• Ensured all tensors used in RGF and related calculations are allocated on or moved to the correct device (rgf_device), including Hamiltonian blocks, self-energies, and intermediate matrices in both batched and non-batched scenarios. [1] [2] [3] [4] [5] [6]
• Updated tensor operations and file I/O to handle device transfers, including moving tensors to CPU before saving with .numpy() and ensuring device consistency during block construction and property calculations. [1] [2] [3] [4] [5] [6] [7]

Documentation and argument updates:

• Clarified in comments and argument documentation that Hamiltonian initialization and self-energy calculations always run on CPU, while only the RGF step uses the specified device. Added a new rgf_device argument to the CLI and configuration. [1] [2]

These changes enable flexible GPU acceleration for the RGF step, improve code clarity, and ensure robust and error-free device handling across the codebase.

TODO:
For now, in RGF, all sub-blocks Hamiltonian and Overlap are transfered to GPU at once. For large systems, OOM would happen. More detailed optimizaiton should be implemented.