benchmarks: add prost bytes-feature variant + MediaFrame for zero-copy comparison

[iainmcgin]

Adds a benchmark variant that enables prost's bytes::Bytes substitution (prost-build::Config::bytes(["."])) and a bytes-heavy MediaFrame message so the feature has something to work with. Motivated by issue #56 ("Clarify zero-copy semantics") — the commenter correctly pointed out that prost has a similar-in-spirit feature and asked for a fair comparison; this PR puts one in the README.

What's new

• benchmarks/prost-bytes/ — mirrors benchmarks/prost/ but builds with .bytes(["."]) and decodes from bytes::Bytes input (so Bytes::copy_to_bytes is the ref-count slice, not a copy). Only decode / merge are measured — the substitution does not affect the encode path, so those benches would be redundant with the existing prost crate.
• bench.MediaFrame — primary bytes body (1–10 KB) + repeated bytes chunks (2–6 × 0.2–2 KB) + map<string, bytes> attachments (0–4 × 50–500 B). Plumbed through gen-datasets, bench-buffa (owned + view + json), bench-prost, bench-prost-bytes.
• task bench-prost-bytes and task bench-cross-prost-bytes; task bench-cross now runs all three variants and produces benchmarks/results/prost-bytes.json. benchmarks/charts/generate.py gains a prost (bytes) series and a MediaFrame row; the README Performance section is updated.
• Small cleanup in gen-datasets: random_bytes now uses rng.fill_bytes instead of an elementwise rng.random() loop.

Headline numbers (MiB/s, task bench-cross, Docker)

Binary decode

Message	buffa	buffa (view)	prost	prost (bytes)
ApiResponse	862	1,475	756	676
LogRecord	722	1,984	712	676
AnalyticsEvent	199	320	254	194
GoogleMessage1	1,014	1,341	956	931
MediaFrame	16,816	73,004	9,648	23,516

On the four non-bytes messages, prost (bytes) tracks default prost within noise — the substitution only affects proto bytes fields, and these messages have none. On MediaFrame it's ~2.4× default prost, confirming the feature lands when it has bytes fields to work with. buffa (view) stays well ahead of both prost variants because it borrows strings, messages, and map/repeated scaffolding from the input buffer too, not just bytes payloads.

perf stat evidence (MediaFrame decode, native)

Native run with perf stat -e task-clock,cycles,instructions,L1-dcache-load-misses,branch-misses around --measurement-time 6 --sample-size 10; per-decoded-message rates (non-hermetic, ±few-% run-to-run):

variant	GiB/s	cyc/msg	ins/msg	L1-miss/msg	IPC
prost	9.1	3,929	10,470	113	2.67
buffa owned	16.6	2,101	5,679	91	2.70
prost (bytes)	21.9	1,632	4,625	5	2.83
buffa (view)	69.8	508	1,675	4	3.30

prost (bytes) closes the cache-miss gap (113 → 5 L1 misses/msg — allocator traffic from cloning bytes payloads is gone). It still runs ~2.8× the instructions of views because String::from_utf8 / HashMap::insert / Vec::push work is still happening for the non-bytes fields (frame_id, content_type, map keys, repeated scaffolding). Views skip that scaffolding — strings stay &'a str, attachments is a MapView, chunks is a RepeatedView.

Files touched

• new benchmarks/prost-bytes/ + benchmarks/Dockerfile.bench-prost-bytes
• benchmarks/proto/bench_messages.proto — MediaFrame
• benchmarks/gen-datasets/src/main.rs — gen_media_frame, random_bytes → fill_bytes
• benchmarks/buffa/benches/protobuf.rs — MediaFrame owned + view + json
• benchmarks/prost/benches/protobuf.rs — MediaFrame decode + json
• benchmarks/charts/generate.py — prost (bytes) series, MediaFrame row
• Cargo.toml, Taskfile.yml, README.md, chart SVGs, tables.md

Notes for reviewers

• The encode-side numbers for prost (bytes) are deliberately absent (same codepath as default prost — see the module docstring on benchmarks/prost-bytes/src/lib.rs).
• protobuf-v4 and Go don't have a MediaFrame row — those suites have a fixed set of messages we didn't extend for this experiment. bench-cross ignores missing rows cleanly.
• gen-datasets' module boilerplate gained #[allow(clippy::enum_variant_names, clippy::upper_case_acronyms, ...)] attributes: on current main the generated log_record::Severity enum variants (UNSPECIFIED/DEBUG/...) and the oneof Value* names trip those lints. The attributes match what's applied to generated code elsewhere in the workspace.

[asacamano]

Is this a generated file?

[iainmcgin]

oh whoops yeah let me check if I can deterministically produce these at the start of runs instead of committing them

[iainmcgin]

actually will keep it this way for now, and replace with a pre-generation step via taskfile later