fix(server): correct MessageData annotations and materialize dynamic fields (#49)

[ff225]

Summary

• MessageData's core field annotations didn't match runtime types: device_id/message_id are str (decoded from bytes in inference_protocol.py), message_content is a dict, and timestamp is a float (<d in the wire format), not int/str as declared.
• The post-decode attributes (received_timestamp, avg_speed, latency, synthetic_latency, payload_size, offloading_layer_index, layer_output, device_layers_inference_time, device_cpu_percent) were unannotated class variables, so @dataclass never turned them into real fields even though every caller treats them as instance state.

Scoped to the typing-correctness part of #49 (annotations + materializing the dynamic attributes as fields). Left get_latency's string-coercing signature and the broader "adopt TypedDict/slots for recurring config structures" follow-up out of scope — no single obvious target and no behavior change needed there.

Note: payload stays str — the only construction site (inference_protocol.py) always passes "" and no code reads message_data.payload, so there's no evidence it's ever actually bytes today; flagging in case that's stale intent from before the protocol refactor.

Test plan

• tests/unit/test_message_data.py: new tests assert the core field annotations match runtime types, the previously-dynamic attributes are real dataclasses.fields, and they default to None.
• uv run pytest fast suite (217 passed, 1 skipped)
• ruff check / ruff format --check on changed files