tracing: add the v2 shared-ring chunk protocol

The producer-local shared ring from RFC 0014: fixed-size chunks, many
writers, one reader, and no waiting for a writer. The protocol document
lands with it; the section numbers in the comments refer to that file.

- One atomic uint32_t per chunk carries everything a handoff needs, so
  that word's modification order decides every race. 32 bits because a
  64-bit atomic is not lock-free on every target we support.
- Five states: FreeForWrap, Acquired, Complete, RewriteRequested,
  Acknowledged. The free word holds the 29-bit wrap count of its
  traversal, so "vacant" and "vacant for your reservation" are one CAS.
  No tombstones, and a writer that slept cannot publish behind the
  reader.
- Only the reader writes a free word, always deriving the wrap from the
  position it is resolving.
- Payload grows up from the header, fixed-width size entries grow down
  from the end. Nothing is backfilled, so the reader takes a committed
  prefix and leaves the writer its suffix.
- SharedRingBuffer owns every state-word transition; RingWriter and
  ChunkReader sit on top. The reader validates its own copy of a chunk,
  never the mapping.
- Backpressure is writer-waits-on-read_pos, Linux and Android only. A
  call that burned a reservation returns kNoChunkAvailable instead of
  waiting: those holes are the reader's to resolve and nothing has asked
  it to run yet.

Not wired up: no SDK writer, no bridge, no muxer, no flag.

Tests: the ABI words and worked chunk from the document, geometry and
wrap arithmetic across the cursor rollover, the three two-party races on
two threads through a step sequencer, scraping and relocation, malformed
and unknown input, the futex paths and their errno policy, the control
block layout, a forged write_pos, and MPSC stress that checks
exactly-once and per-writer order rather than throughput.

Bug: 536851377