[Kinflou]

Stage 1 of runtime#29: benchmark gates. Was to be stacked on #32, which has merged, so this targets main.

• Shared definitions, runtime/conformance/benches/: five format shapes (small scalar, string-heavy, nested, 1000-element array, enum-heavy) in the vectors' value and type encoding, framing payloads of 16 B to 64 KB, the handshake frame and one connection step, with a README of the format and the bench ids. A conformance test checks each shape round-trips through Wire and decode_typed in all three formats.
• iai-callgrind benches, runtime/core/benches/runtime.rs (58 benchmarks, iai-callgrind pinned to 0.16.1): encode and decode through serde-derive mirrors of each shape, framing encode and decode for datagram@1, datagram@2 and jsonrpc@1, handshake encode, decode and check, and State::encode_request then receive. Inputs come from setup functions. A test asserts each mirror encodes to exactly its definition's bytes in every format.
• Gate, .github/workflows/bench.yml: benches the base in a git worktree with --save-baseline=base, then the head with --baseline=base --callgrind-limits=ir=1%. A base with no benches (this PR) records only. The comparison goes to the job summary, and rust.bench.json is uploaded as an artifact (not committed).
• Noise and threshold:
  • The same commit, run three times locally, differed by 0 instructions on all 58 benches. In CI, a no-op change compared against this branch was also 58 of 58 unchanged.
  • With the default codegen units, slowing only the handshake encode also moved jsonrpc@1 encode by up to +2.6%, with no change to its code. [profile.bench] codegen-units = 1 removes that.
  • The limit is ir=1%.
• Gate demo (throwaway [demo, do not merge] bench gate check #36, closed): an extra 20-iteration loop in HandshakeFrame::encode failed the job with handshake/encode 336 → 357 (+6.25%); the other 57 benches were unchanged.
• Base and head build into separate target directories. In a shared one, Cargo reused the base's build for the head: the path packages hash the same from either checkout, and the head's sources look older. The first demo run passed because of this. They still share iai-callgrind's output directory for the baseline.
• No --regression-fail-fast: it stops at the first regression and prints no summary. Without it, every bench is compared and the run still exits non-zero.