Add the analyze_async_overlap CLI subcommand to measure how much of each asynchronous XLA operation (collectives, async copies, send/recv) is hidden behind compute on the same device, and how much time multiple engines on the same TPU chip (TensorCore and SparseCore) spend busy concurrently:

[copybara-service]

Add the analyze_async_overlap CLI subcommand to measure how much of each asynchronous XLA operation (collectives, async copies, send/recv) is hidden behind compute on the same device, and how much time multiple engines on the same TPU chip (TensorCore and SparseCore) spend busy concurrently:

• cli/internal/async_overlap.py: Pure-Python interval analysis over device TimelineEvents (XLA Ops, Async XLA Ops, Sparse Core Ops). Uses Async XLA Ops windows when present and falls back to FIFO <op>-start / <op>-done pairing on the op lines when absent. Classifies each async window as hidden (covered fraction >= hidden_threshold, default 0.9), exposed (<= exposed_threshold, default 0.1), or partial, and computes pairwise concurrent busy time across engines per TPU chip.
• cli/internal/{google,oss}/xplane_tools.py: Adds iter_timeline_events to extract device timeline events with absolute picosecond start times across both XPlane backends.
• cli/tools/analyze_async_overlap_tool.py: CLI/MCP tool wrapper registered as analyze_async_overlap in xprof_cli.py and server.py.
• skills/xprof/: Documents analyze_async_overlap in references/analyze_async_overlap.md and SKILL.md.