Use the standalone pycuampcor package for dense offsets and offsets product

[lijun99]

This PR replaces isce3's internal copies of pycuampcor with the standalone pycuampcor package (https://github.com/earthdef/cuAmpcor, release 2.2.0): one code base with a CUDA backend (PyCuAmpcor) and a multi-threaded (OpenMP) CPU backend (PyCPUAmpcor). #410 fixes the internal copies, so that the switch can be checked bit for bit.

Changes

• Internal pycuampcor removed: the sources and bindings under cxx/isce3/{,cuda/}matchtemplate/pycuampcor are removed. For compatibility, isce3.matchtemplate and isce3.cuda.matchtemplate re-export PyCPUAmpcor and PyCuAmpcor from the package.
• dense_offsets and offsets_product use pycuampcor, and offsets_product can now also run on the CPU.
• offsets_product processes its layers together (runAmpcorLayers): each image chunk is loaded once for all layers. With an older pycuampcor, the layers run one at a time as before.
• The reference RSLC is read directly from HDF5 (HDF5:<file>:<dataset>), so the ENVI copy is no longer made. It is still made for complex32 data or an older pycuampcor without HDF5 support.
• Automatic batch size: windows_batch_range/windows_batch_azimuth may be empty or 0 (now the default), meaning (SM/4) x 8 windows on the GPU (e.g., 20 x 8 on a V100, 27 x 8 on an A100) and 1 x 1 on the CPU.
• New option cross_correlation_workflow (two_pass, the default, or one_pass) for dense_offsets and offsets_product.
• Tests and environment: the ampcor tests use pycuampcor; environment.yml requires pycuampcor>=2.2.

Results

• Same results as develop with Fix bugs in the internal pycuampcor: zoom-in shift at the negative search limit, reduction races, memory leak #410: offsets_product on two NISAR frames (Jakobshavn Isbrae, track 170D, frame 53; Pine Island Glacier, track 149A, frame 140), reprocessed with the SDS ROFF runconfigs. All 173 datasets of the products are bitwise identical between this PR (layers together, RSLC from HDF5) and develop with Fix bugs in the internal pycuampcor: zoom-in shift at the negative search limit, reduction races, memory leak #410 (layers one at a time, ENVI copy), on a V100. Processing the layers together or one at a time also gives bitwise identical products, on a V100 and an RTX PRO 6000.
• Against develop as it is, the only differences are those explained in Fix bugs in the internal pycuampcor: zoom-in shift at the negative search limit, reduction races, memory leak #410.
• Speed, offsets_product (the three layers, compared with develop and the SDS batch of 5 x 10): 2.4x faster on Jakobshavn on an RTX PRO 6000 (752 s to 309 s). On a V100, 2.25x faster on Jakobshavn (1889 s to 840 s) and 2.0x on Pine Island (1716 s to 857 s). The gains come from the automatic batch, the layers processed together, buffers allocated once, and no ENVI copy. Peak resident memory drops from about 66 GB to 3-4 GB.
• Speed, dense_offsets with the SDS GUNW settings: 5-9x faster on the GPU (including develop's ENVI copy), and about 10x faster on 16 CPU threads than develop's single-threaded CPU code.

Dependency and CI

• Temporary channel: pycuampcor is submitted to conda-forge (conda-forge/staged-recipes#34997) but not yet available there. Until then, environment.yml adds the anaconda.org channel lijun99. It holds the same recipe, built for the conda-forge variants by a GitHub Actions workflow in earthdef/cuAmpcor: cpu, and CUDA 12.9, 13.0 and 13.4, for linux-64/aarch64 and osx-64/arm64, Python 3.11-3.14. The channel can be removed once the package is on conda-forge.
• Environments solve for every CI job: with this channel, the environment solves for each job: Linux latest, minimum Python, minimum C++ standard, macOS, and the GPU job with the CUDA toolkit from the nvidia channel. The GPU job gets a CUDA build of pycuampcor.
• Build dependencies: pycuampcor needs HDF5 and zlib at build time, which isce3's environment already has.

Testing

The CI jobs were reproduced in fresh environments created from this branch's environment.yml, so pycuampcor was installed from the lijun99 channel. Each build was installed and the CI tests run serially, as in the workflows (ctest -E ".*(stage_dem|pybind\.unwrap\.phass)").

• GPU job (RTX PRO 6000; the CUDA toolkit installed with conda install -c nvidia cuda, CUDA 13.3; WITH_CUDA, RelWithDebInfo): the environment gets the CUDA build of pycuampcor (cuda130).
• CPU job (no GPU driver; CPU-only build, Debug): the environment gets the cpu build of pycuampcor.

With #411, all tests pass in both jobs.