Add GPU-accelerated Bluestein FFT for Fresnel propagation

[xiaziyna]

• New module: bluestein_fft_gpu.py — PyTorch/cuFFT backend for chunked Bluestein zoom FFT, achieving 15x speedup over CPU (2.5 min vs 60 min on 99k x 99k starshade mask, RTX 3090)
• CPU v2 optimizations to bluestein_fft.py: dict-based chirp cache, pre-allocated buffer reuse, in-place multiplies, scipy.fft workers=-1
• GPU auto-detection via use_gpu parameter on FresnelSingle, threaded through simulate_field and propagator.gen_pupil_field
• Optional dependency: pip install pystarshade[gpu] for torch
• Fix: N_chunk parameter was ignored (hardcoded to 4) in nchunk_zoom_fresnel_single_fft and simulate_field
• Tests: 6 new GPU tests + all 38 existing CPU tests pass