NIMPA: High-throughput neuroimage processing and analysis

add N-dimensional conv_separable() GPU (fallback: CPU) function (#15) use https://github.com/AMYPAD/cuvec instead of numpy arrays (avoids all unnecessary memory copies) (#16) allow arbitrary 3D input dimensions (auto-pad) add tests fix alloc bug (cuvec>2.1.2, https://github.com/python/cpython/pull/24283) build updates ensure can still build without CUDA cmake: allow linking shared libs for external builds framework: misc test & build updates (#15, #14, #13) fix cuda compute capability>=3.5 auto-detection misc minor metadata update