Add reproducible Cloud TPU v5e setup workflow

[charlesmartin14]

Summary

• add baseline/nanogpt_one_head/setup_tpu_v5e.sh
• capture the known-good PyTorch 2.6 / PyTorch-XLA 2.6 installation sequence
• repair stale UNKNOWN-0.0.0 packaging artifacts and old pip/setuptools/wheel
• persist TPU environment variables across SSH sessions without hard-coded home paths
• support explicit ephemeral or persistent storage
• verify raw XLA devices and rg-onehead-env --device auto
• optionally prepare and run a small non-scientific AdamW TPU smoke benchmark
• update TPU.md with Flex-Start provisioning, SSH, setup, diagnostics, and cleanup

Validation

• bash -n baseline/nanogpt_one_head/setup_tpu_v5e.sh
• verified --help output
• verified the committed Git blob SHA matches the locally validated script
• verified the utility and documentation satisfy the repository's no-home-path and no-wrapper-directory invariants
• installation sequence is grounded in the successful v5litepod-4 session: torch 2.6.0+cu124, torch_xla 2.6.0, and four visible XLA devices