ci: pin torch in build/test requirements to match the wheel range

Both `env/build_requirements.txt` and `env/test_requirements.txt`
left `torch` unpinned, so `uv pip install --extra-index-url
https://download.pytorch.org/whl/cu130 ...` resolved it to whatever
was newest on the CUDA index.

That worked while the newest cu130 torch matched `pyproject.toml`'s
declared wheel range (`torch>=2.8,<2.12`). Once torch `2.12.0+cu130`
was published, the build env started installing 2.12.0+cu130 (so
fvdb's `.so` was compiled against the 2.12 ABI, which includes
`at::TensorBase::const_data_ptr<int8_t>`), but the *test* env's
install of the resulting wheel hit pyproject's `<2.12` ceiling and
pip downgraded torch to the CPU-only `2.11.0` from PyPI (which is
missing those CUDA-side template symbols). The CUDA 13.0 unit-test
pipeline then surfaces:

    ImportError: .../libfvdb.so: undefined symbol:
      _ZNK2at10TensorBase14const_data_ptrIaEEPKT_v

at `import fvdb` time. Every CUDA 13.0 PR since 2026-05-16 has been
red on this; the CUDA 12.8 pipeline isn't affected (its index still
serves a wheel in-range).

Pinning `torch>=2.10,<2.12` in both files brings build and test
back onto the same wheel and reflects the wheel's declared range
directly. A follow-up PR can drop the upper bound once the rest of
fvdb has been verified against torch 2.12.

Signed-off-by: Francis Williams <francis@fwilliams.info>

Author: fwilliams

env/build_requirements.txt

@@ -3,4 +3,14 @@ scikit_build_core
 wheel
 ninja
 numpy
-torch
+# Pin torch to the range pyproject.toml declares for the wheel
+# (`torch>=2.8,<2.12`). Without an explicit pin here `uv pip install`
+# resolves `torch` to the latest version on the CUDA index; once
+# torch 2.12 was published, the build env started installing
+# 2.12.0+cu130, fvdb's `.so` ended up compiled against the 2.12 ABI,
+# and then the test step's install of the resulting wheel hit
+# pyproject's `<2.12` ceiling and downgraded torch to the CPU-only
+# 2.11.0 from PyPI -- which is missing those CUDA-side template
+# symbols and surfaces as `ImportError: undefined symbol` at
+# `import fvdb` time on the CUDA 13.0 pipeline.
+torch>=2.10,<2.12

env/test_requirements.txt

@@ -4,7 +4,10 @@ wheel
 ninja
 tqdm
 numpy
-torch
+# Pin must match env/build_requirements.txt so the test env loads
+# the same torch ABI the wheel was compiled against. See the longer
+# comment in build_requirements.txt for the failure mode this avoids.
+torch>=2.10,<2.12
 GitPython
 pandas
 pytest