Refactor RGF kernel and Add no_grad decorator

[AsymmetryChou]

This pull request introduces a significant performance and usability improvement to the recursive Green's function (RGF) calculation in dpnegf/negf/recursive_green_cal.py by adding support for a "stacked" mode for uniform block matrices. When all device blocks share the same shape, the code now auto-detects this and processes the calculation in a fused, batched way, reducing Python overhead and improving efficiency. The change is backward-compatible: outputs are unstacked to lists for downstream compatibility. Additional refactoring and docstring improvements clarify the new behavior and streamline the code.

The most important changes are:

Performance and Algorithm Improvements

• Added a stacked mode to recursive_gf_cal that processes uniform block matrices as stacked tensors, enabling a single fused operation and reusing cached identity matrices for improved efficiency. The function auto-detects when this mode can be used and falls back to the legacy list-based path otherwise. [1] [2] [3] [4] [5] [6]
• Updated the wrapper recursive_gf to detect when all blocks are uniform, stack them for the kernel call, and unbind outputs to maintain the previous API contract. [1] [2]

Code Quality and Documentation

• Improved and expanded docstrings for both recursive_gf_cal and recursive_gf to document the new stacked mode, its requirements, and its effect on performance and outputs. [1] [2]
• Refactored code for clarity: removed unnecessary else branches, hoisted repeated matrix multiplications to variables, and added comments explaining the uniform/stacked path. [1] [2]

Minor Fixes and Cleanups

• Added @torch.no_grad() to the NEGF class constructor for improved efficiency and removed an unnecessary context manager. [1] [2]

These changes collectively improve both the performance and maintainability of the RGF code, especially for large systems with uniform block structure.