Replace the Buck stub tests with one shared harness

Summary:
The torch and numpy stub packages each carried their own near-duplicate
`run_pyrefly.py`, and the two had drifted: numpy could run Pyrefly out of Buck
but not from a release build, torch could do the reverse. Adding a third package
for jax would have meant a third copy and a third set of gaps. The runtime side
had the same problem from the other direction, with the Buck targets and the
GitHub workflow building their environments in completely different ways, so a
green run in one said little about the other.

Factor the parts that do not vary between packages into `shape_testing.py`: a
`Suite` describing a named group of test files, resolution of the Pyrefly binary
and of the virtualenv, and the two runners. A package now supplies only its
`suites.py` plus thin entry points, so jax will be additive rather than another
copy.

Delete the Buck stub test targets rather than port them. They were a second,
divergent way to run the same checks, and two of them built torch from source to
do it. Everything now goes through one runner, so `test.py`, the internal
Sandcastle job and the GitHub job differ only in where the Pyrefly binary comes
from; `--buck` selects it in an internal checkout. The stub `python_library` and
search-path targets stay, because they are how internal tooling and the porting
skills point Pyrefly at the stubs.

Separate bootstrapping from running, which is the main design constraint here.
`bootstrap_venv.py` is the only thing that touches the network; the runners never
create anything and fail with a message naming the fix instead. That matters
because on a Meta machine the download needs fwdproxy, and a sandboxed agent
cannot do it at all without being relaunched with `--secure-internet-mode`.
Keeping the two apart means everything except bootstrap works offline, and the
one step that cannot is reported clearly rather than hanging. One virtualenv at
`~/.tensor-shapes-venv` serves all the libraries, since the runtime tests only
build arrays and read shapes.

Type checking needs no virtualenv at all: Pyrefly resolves the stubs through
`--search-path` and never imports the real library. `run_tests.py --static-only`
exposes that, and the missing-virtualenv error advertises it, because it is the
usual mode when working on Pyrefly itself rather than on the stubs.

Pin the libraries, and pin them to CPU-only builds. These tests assert the shapes
the libraries actually produce, so an unpinned minor bump can turn a correct stub
red. Shape checking never runs a kernel, so the CUDA stack is pure cost: the
resulting virtualenv is 1.3G against 5.1G for a CUDA one.

Run everything as a single job on both CI systems -- `fbcode//pyrefly:oss`'s
sibling `:tensor_shapes` internally, one set of steps on GitHub -- rather than
one job per library. The whole suite takes about 25 seconds, so splitting it
would cost more in scheduling than it catches. The internal job installs `uv`
when the image does not provide it, the way the neighbouring continuous jobs
bootstrap their own tooling.

Two behaviors are preserved deliberately. `GRADUAL_SHAPE_RUNTIME_TESTS` is now
honored by the shared runner: a test whose static shape is gradual may use plain
runtime assertions, because `assert_shape` still demands an exact static shape.
And the two numpy-behavior tests from the deleted `numpy_runtime_test` wrapper
move into `test/test_indexing.py`, where they keep asserting that NumPy rejects
float array indices and mismatched paired index lengths at runtime. Pyrefly does
not reject either statically, so they carry no inline error expectation and a
comment records the gap.

On GitHub, type checking runs on the whole matrix and only the runtime half is
Ubuntu-only. Type checking needs no virtualenv, and the runners carry
Windows-specific interpreter and `.exe` path handling that nothing else would
exercise, so restricting it would leave that code untested on the only platform
it exists for. The runtime half is genuinely platform independent, so installing
torch and jax three times to assert the same shapes would be waste.

Differential Revision: D117306584