tpu-perf

Systematic analysis of TPU pretraining efficiency. profile-anatomy: schema + reference scripts for xplane.pb / trace.json.gz. comm-analysis: per-axis NCCL bus BW, peak ICI utilization, and per-step compute/comm overlap. compute-breakdown: source-line aggregation, layer scoping, non-compute audit, v7x roofline shortfall. memory-profile: HBM peak occupancy, alive-buffer attribution, lifetime/shape/tf_op/parent_jit rollups.