Prevent macOS Server worker stack overflow

[yyjeqhc]

Summary

Fix a macOS-only WebCodex Server crash observed during real OpenAI Secure MCP Tunnel dogfood.

A normal MCP request could reach the local Server and abort a Tokio worker with stack overflow. The same request succeeded when RUST_MIN_STACK=8388608 was supplied diagnostically.

This change makes the webcodex-server binary build its Tokio multi-thread runtime explicitly and uses an 8 MiB worker stack on macOS only. Linux and Windows retain Tokio's default worker stack behavior.

Evidence

• reproduced the crash twice on mini/macOS during real webcodex share --tunnel openai calls
• diagnostic RUST_MIN_STACK=8388608 removed the crash
• macOS runtime regression verifies the configured worker stack is at least 8 MiB
• cargo fmt -- --check passed
• focused Server compilation passed
• macOS arm64 build of WebCodex/Runner/CLI passed
• with RUST_MIN_STACK explicitly unset, live share --tunnel openai reached /healthz=200 and /readyz=200
• direct MCP initialize and tools/list both returned HTTP 200 and the Server remained alive
• real ChatGPT Secure Tunnel tool invocation passed after the stack adjustment

The branch was rebased cleanly onto current main; the intervening #162 change is Windows Runner-only and does not overlap this fix.