tracing: add the v2 shared-ring chunk protocol

Add the producer-local shared ring used by tracing v2. This change contains
the shared-memory ABI, one writer per TraceWriter, the single reader, and the
protocol tests. SDK and service wiring follow in separate changes.

- Define a 64-byte ring header:
  - an always-lock-free atomic64 packs write_pos in the high half and read_pos
    in the low half;
  - num_writers_waiting avoids a wake syscall when no writer is asleep;
  - a zero-filled mapping is a valid empty ring.

- Define one atomic32 state word at the start of each chunk:
  - Free(wrap) exposes the chunk to one ring traversal;
  - BeingWritten identifies the writer and its published fragment prefix;
  - Complete says the writer has finished touching the chunk, although it may
    take the chunk back to append more fragments;
  - RewriteRequested says the reader took the published prefix and the writer
    must move any newer suffix;
  - RewriteAcknowledged says the writer released the old chunk and the reader
    may return it to Free.

- Reserve uint32 positions in FIFO order and map them onto a power-of-two
  number of physical chunks. Free carries a 16-bit wrap count so a delayed
  writer cannot claim the same chunk during a later traversal.

- Keep ownership rules local to the chunk word:
  - one reservation permits one exact claim attempt;
  - only the reader writes Free;
  - the reader never waits for a writer;
  - a lost reader CAS leaves read_pos unchanged and retries the position in a
    later drain.

- Store target BufferID after the state word. Payload fragments grow towards
  higher addresses while their sizes grow down from the end of the chunk as
  shortest-form protobuf varints. Fragment counts publish an append-only
  prefix. If the reader reaches an active writer, it copies that prefix and
  asks the writer to move only the unpublished suffix.

- Keep the initial 256-byte minimum chunk size and require a four-byte-aligned
  chunk stride. Check ring geometry and allocation arithmetic before mapping
  memory. Require atomic32 and atomic64 to be always lock-free.

- Honor BufferExhaustedPolicy in SharedRingBufferWriter. Linux and Android
  writers can wait on the read_pos half of rw_positions with a futex. Other
  platforms report that blocking is unavailable instead of spinning. The
  waiter hint and its seq_cst fence pair prevent a missed wake across the two
  atomic words.

- Test ABI encoding, position and wrap rollover, malformed chunk input, every
  state transition, failed claims, prefix copying, suffix relocation, futex
  wake races, and concurrent multi-writer stress.

Bug: 536851377