Upstream tokenspeed: msgspec/msgpack IPC mode for io_struct

[slin1237]

Part of #2000; gates #2006.

TokenSpeed's frontend↔scheduler IPC is sglang topology (EngineCoreClient PUSH/PULL over PortArgs) but tokenspeed/runtime/engine/io_struct.py is plain-dataclass, pickle-only — no msgpack mode, so no non-Python frontend can speak it.

Upstream PR to lightseekorg/tokenspeed, using sglang 0.5.16's io_struct migration as the template:

• Port io_struct dataclasses to kw-only msgspec Structs (tag=True, array_like=True base, matching sglang's BaseReq pattern)
• sock_send/sock_recv helpers honoring a pickle/msgpack env toggle (pickle stays default, like SGLANG_USE_PICKLE_IPC)
• Adopt typed msgspec multimodal structs from day one (the Upstream sglang: typed msgspec multimodal structs for msgpack IPC #2004 design) — skip the PickleWrapper intermediate step
• Update AsyncLLM/EngineCoreClient/scheduler send-recv paths to the helpers
• Tests: round-trip parity pickle vs msgpack for generate/embedding/abort/output structs

Lessons from the sglang 0.5.16 consumer-side port (smg #1970) worth baking in: slotted structs turn phantom-attribute writes into hard errors (audit setattr sites); msgspec reports missing required args one at a time (make required fields deliberate); array("q") token-id containers.