[copybara-service]

Run the PyTorch H2D/D2H perf gate in GitHub CI

The h2d_d2h_gating workload ran with the default --framework=jax, so the torch
floors were never checked. This makes the torch gate mirror the JAX one: it
runs on the same runner, enforces floors recorded there, and is a required
check.

• Add ci/run_in_torch_env.py, which sets up an open-source torch environment
  and runs a script in it: installs torch, torch_tpu and libtpu 0.0.50, builds
  the torch extension with build.sh torch (it can't come from bazel run),
  links it to torch_tpu's glue library, and configures libtpu for a
  multi-process slice in a GKE container.
• Add the h2d_d2h_gating_torch workload (required check in copy.bara.sky) and
  the h2d_d2h_record_torch workload; pass --framework=jax to the JAX workloads
  explicitly and move the torch_* metrics to the torch workload.
• Re-record the torch floors on the CI runner (--record --iters=500).
• Use the 2x4 v5e-8 topology in perf_core's open-source build.
• Document that floors are recorded only on the CI runner.

[google-cla]

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.