fix(gateway): add accept-loop backoff to prevent CPU spin | 修复(gateway): 为 accept 循环增加退避，避免 CPU 空转

[manelsen]

Fixes #851

Summary

EN:

• Hardened the gateway accept loop to apply bounded backoff on non-WouldBlock accept errors, preventing tight CPU spin under repeated transient failures.
• Added explicit backoff helpers/constants (ACCEPT_ERROR_BACKOFF_MAX_MS, log interval control) so retry behavior is deterministic and observable.
• Added regression coverage in src/gateway.zig for the backoff progression and cap behavior.

ZH:

• 加固了 gateway 的 accept 循环：在非 WouldBlock 的 accept 错误下应用有上限的退避，避免瞬时错误反复发生时出现 CPU 空转。
• 增加了明确的退避辅助逻辑与常量（ACCEPT_ERROR_BACKOFF_MAX_MS、日志节流间隔），使重试行为可预测且可观测。
• 在 src/gateway.zig 增加回归测试，覆盖退避增长与上限行为。

Validation

• zig build test --summary all

Notes

• This change is intentionally scoped to gateway accept-loop error handling and does not alter successful connection flow.

[rijuyuezhu]

Met the same problem on Arch Linux with nullclaw of latest dev version

[mark-os]

Posted a root cause analysis on #851 — the WouldBlock branch is unreachable on Zig 0.16 because Io/Threaded.zig maps EAGAIN to error.Unexpected via errnoBug(). This PR's backoff on the else branch is the right short-term fix, but the compat shim should probably remap Unexpected back to WouldBlock when the socket is known non-blocking.