Time and memory efficient parallelization

[antelmor]

Summary

Introduces a new parallelization scheme designed to improve both runtime performance and memory efficiency for computing Green’s functions and the $A_{ij}$ exchange tensor.

The core of this update is a new lightweight execution object, GreenRuntime, whose sole responsibility is to perform the heavy numerical work while minimizing memory footprint and inter-process communication overhead.

---

Key Design Changes

1. GreenRuntime: lightweight compute runner

• Introduces a dedicated GreenRuntime object that encapsulates only the minimal information required to compute Green’s functions and the $A_{ij}$ tensor.
• Heavy data structures from initialization (e.g. full tight-binding Hamiltonians) are not stored, unlike in ExchangeNCL or TBGreen.
• Avoids large intermediate dictionaries in favor of contiguous NumPy arrays, improving cache efficiency and serialization speed.

2. Efficient multiprocessing via np.memmap

• Large arrays are stored using np.memmap, ensuring:
  • Minimal memory duplication across processes
  • Reduced inter-process communication overhead
  • Safe scaling to larger systems
• This design enables faster multiprocessing without increasing memory pressure.

3. Cross-platform multiprocessing support

• The new scheme is compatible with both multiprocessing start methods:
  • 'fork' (Linux)
  • 'spawn' (macOS, Windows)
• This ensures portability across operating systems without changes to the execution logic.

4. Thread control via threadpoolctl

• Adds threadpoolctl support to limit the number of threads used by NumPy/SciPy backends.
• Prevents thread oversubscription, which can otherwise degrade performance when combined with multiprocessing.

---

Additional Fixes

• Fixed Python syntax in magnons/io_exchange2.py that previously required Python ≥ 3.11.
• The module is now compatible with older Python versions, restoring usability of the magnon scripts.