Set up TPU benchmark runner

[Tyronita]

Context

The TPU benchmark workflow (.github/workflows/tpu-benchmark.yml) requires a self-hosted runner with a TPU, which we don't currently have. Our Azure VMs are GPU-only (A100, T4).

Options

1. Google Colab — Free TPU v2-8 access. Can run benchmarks manually but not as a persistent CI runner (sessions are ephemeral).
2. GCP Cloud TPU VM — Register as a self-hosted GitHub Actions runner. Requires GCP project with TPU quota.
3. Kaggle notebooks — Free TPU v3-8 access, same ephemeral limitation as Colab.

Current workaround

• Colab notebook for one-off TPU benchmarks (see notebooks/tpu_benchmark.ipynb)
• TPU workflow disabled in CI until a persistent runner is available

TODO

• Decide on TPU access strategy
• Set up persistent TPU runner OR scheduled Colab/Kaggle runs
• Re-enable TPU CI workflow