Drop cu130 from the CUDA build matrix

[shoumikhin]

Problem

PyTorch announced that it is removing CUDA 13.0 (cu130) from its nightly builds, and asked
anyone who pins cu130 to move to cu132. The cu130 test channel is already gone. Linux cu130
nightlies came back on October 3, but only as a short-term hold for one downstream project,
not as a supported channel. ExecuTorch follows PyTorch, so its cu130 wheels lag behind too.

The build matrix still listed cu130. That caused two problems:

1. The matrix keeps building a channel PyTorch is retiring.
2. The daily job that moves the ExecuTorch pin forward needs a version that every listed
   channel has. A lagging cu130 holds the pin back, even when cu132 and cu134 already have
   newer wheels.

Solution

Remove cu130 from the x86 and Arm CUDA lists in the matrix filter. The pin job reads the same
lists, so the pin can move to the newest version that cu132 and cu134 share.

Release wheels follow the same lists, so cu130 release wheels are dropped as well, on every
platform. The C++ tarballs keep cu130, because libtorch still ships it.

Once the pin can move past cu130, anything that installs the pinned ExecuTorch from the cu130
index would break. So the places that still use cu130 move to cu132:

• the docs workflow, which installs the pin on every push to main
• the install-test-ext recipe in the justfile
• the runtime wheel README
• the default PyTorch index in pyproject.toml, with uv.lock regenerated
• the install guides, the contributing guide, and the example install hints

A test now checks that the pin job rejects cu130, so putting it back in the lists fails a test.
The Jetson cu126 path is unchanged.

Testing

• Ran the ExecuTorch pin, pin update, shared runtime workflow, API, and packaging test suites.
  The new cu130 rejection case fails when cu130 is put back in the lists.
• Dry-ran the docs, justfile, and README install commands for Linux. With a pin that only cu132
  and cu134 have, they fail against cu130 and resolve against cu132.
• Regenerated uv.lock on Linux with the uv version CI uses, and checked it with
  uv lock --check.
• Did not run the docs workflow or a release workflow, and did not build on a GPU.

[lanluo-nvidia]

lgtm