perf(client): reduce suspended OneWay state allocations

[SunSi12138]

Summary

Related to #729.

Split the existing internal OneWay async core into plain and streaming state machines after one descriptor check. Public/generated APIs and wrapper/lifecycle ownership stay unchanged.

This production PR contains only two files: the runtime change (+29 net lines) and two descriptor-shape regression cases. Tests verify producer ownership, wire flags, owned cancellation and pending/admission cleanup. Research infrastructure and experimental static-entry API B remain in unmerged Draft #795.

Allocation benefit

Hosted investigation 37945758812, using frozen A/C/B source 6ece2f7a and baseline 553854f5:

• All 216 controlled and 108 tiered-PGO actual generated-invocation samples passed exact payload/frame/stream checks.
• Paired C-minus-A process-allocation medians for suspended 0/1/2 streams: −48.024/−7.986/−8.165 B/op (12 pairs), and −47.957/−8.036/−8.276 B/op (six tiered pairs).
• Every suspended pair improves. Synchronous caller allocations are essentially unchanged.
• All C/A timing confidence intervals include 1. No latency improvement is claimed.
• B has no justified incremental benefit over C, so its added API/generated-runtime compatibility cost is rejected.

Code-size tradeoff

Full-RPC 64-bit scalar/no-stream state shrinks 248→192 inline bytes; one/two-stream states 256/272→248/264 bytes. These are struct sizes, not heap-allocation measurements; the local fixture uses a different 32-bit request.

Aggregate MoveNext IL grows 1,781→2,078 B (+16.7%). In this six-shape fixture, NativeAOT executable size grows 25,072 B (+0.243%), and strict core native code grows 21,551→27,688 B (+28.5%). Individual executed bodies shrink but generic instantiations emit both branches; other contract catalogs can differ.

Validation and evidence

• All 324 full-RPC service-completed cases passed: TCP/SharedMemory, JIT/NativeAOT, 0/1/2 streams, batch1/8, deadlines, cancellation, interceptors, telemetry, endpoints and admission.
• This is one-round regression sanity, not statistical proof of zero regression or E2E acceleration. Median C/A wall ratios are 1.0013 JIT and 1.0010 AOT. Noisy outliers include +30.6% JIT TCP one-stream interceptor and +23.9% AOT SharedMemory two-stream interceptor; same-core B is near A in those cases.
• Hosted environment: Ubuntu24.04 x64, .NET SDK10.0.112/runtime10.0.12.
• Original archive SHA-256: 48c10e5d5244a715251be0bba1ae23ee7de435be96753e191f99adc0e28ecc89. Artifact-only recovery verified that digest and retained original raw rows/disassembly without rerunning measurements: 37950819696, 37951510883.
• Independent review verified 763 recovered files, every raw pair, exact full-RPC completion, state layouts, API identity and native/JIT disassembly. Differential lifecycle checks passed 58 tests on A/C; 25 local C integration checks passed, with local SharedMemory permission-blocked before connection and covered successfully on GitHub.

Original investigation-head CI passed. Independent final review verified that two-file head a90c83db0c1c549e9e1352b1f8cdcc2f42df35b7 preserves the reviewed runtime/test blobs. This PR is Ready for review as requested by the owner; final-head CI continues in the background and is not represented as complete. Large Actions artifacts retain seven days; raw research evidence stays separate in #795, never in this production diff.