Refactoring all test scripts to pytests and fixing CUDA graphs compatibility

[akshaysubr]

This PR includes three major improvements to cuHPX:

1. Test Suite Refactoring - Modernized all tests to use pytest with proper fixtures and dtype parameterization
2. CUDA Graph Compatibility - Added support for capturing SHTCUDA/iSHTCUDA operations in CUDA graphs
3. Bug Fix - Fixed segmentation fault caused by dangling CUDA stream references

[Copilot]

Pull request overview

This PR modernizes the cuHPX test suite and adds critical CUDA graph support for production workloads.

Changes:

• Refactored all test scripts from interactive input-based scripts to pytest-based unit tests with proper fixtures and parameterization
• Added CUDA graph compatibility by fixing stream handling and pre-allocating buffers in SHTCUDA/iSHTCUDA operations
• Fixed segmentation fault caused by dangling CUDA stream references in HealpixFFT/HealpixIFFT classes

Reviewed changes

Copilot reviewed 20 out of 20 changed files in this pull request and generated 7 comments.

Show a summary per file

File	Description
tests/test_sht_two_stream_overlap_profiling.py	Removed old profiling test (replaced by benchmark script)
tests/test_sht_cuda_stream.py	Converted to pytest with CUDA stream validation tests
tests/test_sht_cuda_batch.py	Converted to pytest with batched operation tests
tests/test_sht_cuda.py	Converted to pytest with CUDA vs PyTorch comparison tests
tests/test_sht_bluestein.py	Converted to pytest with Bluestein algorithm tests
tests/test_regridding.py	Converted to pytest with grid regridding tests
tests/test_harmonic_transform.py	Converted to pytest with SHT/iSHT roundtrip tests
tests/test_grad.py	Converted to pytest with gradient flow tests
tests/test_differentiability.py	Converted to pytest with gradient-based optimization tests
tests/test_data_remapping.py	Converted to pytest with ring/nest remapping tests
tests/test_cuda_graphs.py	New file with comprehensive CUDA graph capture tests
tests/test_batch_remap.py	Converted to pytest with batched remapping tests
tests/conftest.py	New pytest configuration with shared fixtures and tolerance helpers
src/harmonic_transform/hpx_fft.h	Fixed stream handling and added pre-allocated buffers for CUDA graphs
src/harmonic_transform/hpx_fft.cpp	Updated to use pre-allocated buffers and in-place operations
pyproject.toml	Added pytest configuration and test dependencies
examples/wmap_optimization.py	New example demonstrating differentiable SHT on WMAP data
cuhpx/hpx_sht.py	Fixed stream handling and pre-computed weights for CUDA graph compatibility
benchmarks/stream_overlap_profiling.py	New benchmark for profiling stream overlap
benchmarks/sht_profiling.py	New benchmark for profiling SHT operations

Comments suppressed due to low confidence (1)

src/harmonic_transform/hpx_fft.h:1

• The condition n_ < n only triggers reconfiguration when the new batch size is larger. If n becomes smaller (e.g., from 8 to 4), no reconfiguration occurs, but the pre-allocated buffers sized for n_=8 will be reused. While the code uses slicing to handle this (line 77), the comment on line 133 is misleading - it should clarify that buffers are sized for the maximum batch size seen, not just 'when batch grows larger'.

/*

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[xiaopoc]

I ran the pytest suite on my end and observed 3 test failures. It seems that these failures occur with precision checks and are related to numerical tolerances thresholds being set too tightly. Maybe we can try the following to fix?

1. tests/conftest.py (line 86) – Increase atol for float32 implementation comparisons

   # Current
   torch.float32: {"sht": (1e-4, 1e-5), "isht": (1e-3, 1e-2)},
   # Suggested
   torch.float32: {"sht": (1e-4, 2e-05), "isht": (1e-3, 0.1)},

2. tests/test_sht_cuda_batch.py (line 66) – Increase atol for batch vs single comparison

   # Current
   atol = 1e-6 if dtype == torch.float32 else 1e-10
   # Suggested
   atol = 1e-4 if dtype == torch.float32 else 1e-10

[akshaysubr]

@xiaopoc Thanks for the review and independent testing. What GPU, CUDA version, etc were you testing with?

[xiaopoc]

All tests passed. The PR looks good to me, so approved it.