Add isolated v5e hardware probe to NanoGPT TPU port

[charlesmartin14]

Adds python experiment.py probe to measure the real dense model, replicated optimizer/tail state, host memory, per-chip HBM, XLA compilation and step wall times without loading data, constructing HostTable, building the canonical mask or starting WeightWatcher.

The probe retains max_seq_len=262144, runs two synthetic full-softmax training steps at 128 and 49152 tokens per rank with optimizer steps 0/1, releases optimizer/tail state, and runs one 262144-token evaluation forward. Per-rank JSON and a finalized summary record failures and cannot pass without all eight ranks completing full-size evaluation. The launcher stops the process group after a reported failure and uses package metadata instead of importing XLA in the launch grandparent.

The training body, schedule, optimizer, CUDA reference, preflight gate and all qualification flags are unchanged. README requires a green probe before attempting mandatory preflight. Second-step timings explicitly acknowledge possible first-Adam compilation.

Validation: existing TPU-port/reference CPU suite passed (49 passed, 1 local Gloo socket-restriction skip before the final two reporting tests); final probe tests 10 passed; full 302,792,179-parameter dense CPU smoke passed; all 58 CUDA vendor hashes verified. CI runs mandatory Gloo and XLA-CPU checks. No TPU probe, preflight or training was launched.