Fix Python 3.15 validation on Linux, Windows and macOS

[atalman]

What

Makes Python 3.15 / 3.15t binary validation actually work, on every platform.
#8427 put 3.15 in the default matrix for all OSes, which exposed the uv-based
interpreter setup in validate_binaries.sh — previously Linux-only — to Windows
and macOS. Four bugs, all in that one script:

1. Wrong interpreter pinned (all platforms). import torch segfaults with
every published cp315 wheel under CPython 3.15.0b1, which is what the script
provisioned. The crash is pybind11 3.0.4's gil_scoped_acquire teardown in
python_tracer::init(), reached from torch._C._autograd_init():

torch/autograd/__init__.py:653          if not torch._C._autograd_init():
 └ THPAutograd_initExtension            torch/csrc/autograd/init.cpp:761
    └ python_tracer::init()             torch/csrc/autograd/profiler_python.cpp:1637
       └ pybind11::gil_scoped_acquire::~gil_scoped_acquire()
          └ ::dec_ref() -> PyThreadState_Clear -> SIGSEGV

The same wheel imports and runs fine under 3.15.0b4, so the pin moves to b4.
Verified locally on both interpreters, and with the oldest (dev20260724) and
newest (dev20260805) cp315 nightlies.

2. Windows could not activate the venv. uv lays a venv out as Scripts/, not
bin/ — it says so in the log line right above the failure:

Activate with: conda-env-31013667966\Scripts\activate
+ source conda-env-31013667966/bin/activate
validate_binaries.sh: line 327: .../bin/activate: No such file or directory

3. macOS silently validated the wrong Python. The macos-arm64 workaround
prepends ${CONDA_PREFIX}/bin to PATH. On the uv path no conda env is activated,
so CONDA_PREFIX still pointed at base (python 3.14) and shadowed the venv.
The job did not error — it installed and smoke-tested torch under 3.14, i.e. the
3.15 leg was reporting 3.14 results. The prepend is now skipped on the uv path.

4. Cleanup failed after a green run. ENV_NAME is a conda env name on the
conda path but a venv directory on the uv path, so conda env remove -n fails
with EnvironmentLocationNotFound and, under set -e, fails the job after the
tests passed. The uv path now deactivates and rm -rfs the directory.

5. numpy built from source on Windows. PyPI publishes no cp315/cp315t numpy
wheels, so the smoke-test step's pip install numpy --upgrade falls back to a
source build — slow on Linux/macOS, fatal on Windows:

numpy/_core/src/common/simd/simd.h(4): fatal error C1083:
    Cannot open include file: 'stdalign.h': No such file or directory

download.pytorch.org/whl/nightly already carries preview cp315 and cp315t numpy
wheels for every platform including win_amd64, so 3.15 pulls numpy from there
with --only-binary=:all: to forbid a source fallback. Other Python versions
keep using PyPI unchanged.

The conda path for 3.10–3.14t is untouched throughout.

Test plan

Validated end-to-end with a temporary commit that pointed the workflow checkouts
at this branch and forced PR builds to 3.15 (that commit has been removed;
this PR is now the three real commits, one file).

Final run — every 3.15 leg green:

platform	3.15 legs
linux	7/7 — cpu, cuda 12.6 / 13.0 / 13.2, rocm 7.1 / 7.2, xpu
linux-aarch64	4/4 — cpu, cuda 12.6 / 13.0 / 13.2
win	5/5 — cpu, cuda 12.6 / 13.0 / 13.2, xpu
mac-arm64	green

Also: bash -n clean; shellcheck -S warning reports only the pre-existing
line-181 finding; the 11 matrix unit tests pass.

Follow-ups (not in this PR)

• None of the four validate-*-binaries.yml workflows pass test-infra-ref, so
  generate_binary_build_matrix.yml defaults it to main and the *_job.yml
  templates to "". Both the matrix generator and validate_binaries.sh are
  therefore always taken from main — no PR touching this script can test
  itself without a temporary commit like the one used above. Worth fixing
  properly.
• PR-limited runs validate a single Python version, so 3.15t got no CI
  coverage here. It was verified locally on b4 (GIL genuinely disabled, matmul
  correct, and cp315t numpy wheels exist on the nightly index).
• The pybind11 gil_scoped_acquire fragility behind the b1 segfault is real and
  still worth reporting upstream, even though b4 avoids it.

[vercel]

@atalman is attempting to deploy a commit to the Meta Open Source Team on Vercel.

A member of the Team first needs to authorize it.

[huydhn]

LGTM!