Refactor selfEnergy implementation and lead matrices

[AsymmetryChou]

This pull request introduces performance optimizations and code refactoring to the surface Green's function calculation in the NEGF module. The main improvements are the precomputation and reuse of matrix conjugate-transposes, efficient conversion of tensors to NumPy arrays, and a more consistent interface for the surface Green's function core routines. These changes reduce redundant computations and improve compatibility with Numba and multiprocessing.

Performance and Memory Optimizations

• Added the _pack_lead_matrices utility in lead_property.py to precompute and store conj(HLLk.T) and conj(SLLk.T) as h10 and s10, and to convert relevant matrices to C-contiguous np.complex128 arrays. This minimizes repeated torch→numpy conversions and avoids recomputing conjugate-transposes per energy. (dpnegf/negf/lead_property.py dpnegf/negf/lead_property.pyR911-R948)
• Updated the lead k-point precomputation routines to use _pack_lead_matrices, ensuring all workers receive efficiently packed data. (dpnegf/negf/lead_property.py [1] [2]

API and Interface Changes

• Modified the selfEnergy and surface_green functions to accept optional h10 and s10 arguments for the precomputed conjugate-transposes, defaulting to on-the-fly computation if not provided. (dpnegf/negf/surface_green.py [1] [2]
• Updated all internal calls and the Numba/Scipy core routines to use the new arguments and interface, ensuring consistency and backward compatibility. (dpnegf/negf/surface_green.py [1] [2] [3]

Algorithmic Refactoring

• Refactored the Numba and Scipy Lopez-Sancho core implementations to use batched GEMM operations for updating recursion variables, improving cache efficiency and matching the new interface. Also, replaced the convergence test with a slightly cheaper but equivalent probe. (dpnegf/negf/surface_green.py [1] [2]

Compatibility and Reliability

• Ensured all matrix conversions and precomputations are compatible with both direct NumPy arrays and torch tensors, and with multiprocessing (joblib-memmapped) environments. (dpnegf/negf/lead_property.py [1] [2]
• Added a regression-testable Numba core variant with fastmath=False for validation purposes. (dpnegf/negf/surface_green.py dpnegf/negf/surface_green.pyR171-R176)

These changes should yield a noticeable speedup in NEGF calculations, especially in parallel or high-throughput scenarios, by avoiding redundant computation and memory layout conversions.

[AsymmetryChou]

solve #38 in the last commit @Lonya0