Add run-kaggle-tpu skill

[tinycrops]

What changed

Adds a run-kaggle-tpu skill for agent-driven Kaggle TPU script-kernel runs via the kaggle CLI.

The skill includes:

• SKILL.md runbook for scaffold -> push -> poll -> pull TPU workflows
• scripts/driver.sh helper for scaffolding, pushing, polling, pulling, and one-shot runs
• single-process and 8-core torch_xla smoke-test templates
• metadata templates for TPU-enabled Kaggle script kernels
• README with prerequisites, structure, and installation notes

Why

This packages a verified TPU kernel workflow into the existing official Kaggle skills repository instead of creating a duplicate kaggle-skills repo elsewhere.

It also documents the Kaggle TPU multiprocessing gotcha where TPU_PROCESS_ADDRESSES=local and CLOUD_TPU_TASK_ID must be removed before xmp.spawn(..., start_method="fork").

Validation

• bash -n run-kaggle-tpu/scripts/driver.sh
• run-kaggle-tpu/scripts/driver.sh help
• python3 -m py_compile run-kaggle-tpu/scripts/templates/single_probe.py run-kaggle-tpu/scripts/templates/multicore_probe.py
• python3 -m json.tool on both kernel metadata templates
• rg check for leftover local absolute paths/private workspace references