protozero: let tokenized messages outlive the read loop

[LalitMaganti]

ProtoRingBuffer hands out messages pointing into a buffer that the next
BeginWrite() is free to recompact or grow, so a message is only valid
until the following read. Passing one to another thread - what running
trace processor's RPC off the frontend thread needs - is therefore not
expressible, and the fallback is copying every request out, including
32 MB APPEND_TRACE_DATA chunks.

Refcount the buffer and hand out messages as slices holding a reference.
The write path is unchanged - a single non-wrapping buffer, rewound
whenever the read cursor catches up - except that recompacting, growing
and rewinding are now only allowed while nothing points into the buffer.
When something does, the write moves onto another buffer and the last
message releases the old one.

The buffer left behind is picked up again by the following switch, so a
caller retaining one message per write ping-pongs between two buffers
rather than allocating on each. Measured over 2000 messages, buffers
allocated:

msg size retained before after
512B 0 1 1
512B 1 8 2
512B 64 8 2
64K 1 2000 2
64K 64 2000 2000

The last row is a pool deep enough to pin every buffer at once, where
the allocations are the working set rather than waste.

out/linux_clang_release/perfetto_benchmarks
--benchmark_filter='BM_ProtoRingBuffer.*'

[github-actions]

🎨 Perfetto UI Builds

• da89ac5 (2026-08-28 14:38 UTC): https://storage.googleapis.com/perfetto-ci-artifacts/gh-33180630366-1-ui/ui/index.html
• 64c1068 (2026-08-28 15:19 UTC): https://storage.googleapis.com/perfetto-ci-artifacts/gh-33183888369-1-ui/ui/index.html