fix(streaming): prevent SSE idle disconnects and total stream timeouts

[DKalien]

Summary

Fix long-running streaming responses that can be terminated while the upstream model is still working.

Root cause

The upstream main branch had two independent timeout paths:

• The downstream Axum SSE response did not emit keep-alive comment frames, so Codex could hit stream_idle_timeout_ms during periods without model events.
• OpenCodeGoClient used one reqwest client with ClientBuilder::timeout() for both normal and streaming requests. That timeout applies to the complete response, so a valid stream could be cut off after the normal request budget.

Changes

• Add a 15-second downstream SSE keep-alive.
• Split normal and streaming reqwest clients:
  • normal requests keep the total request timeout;
  • streaming requests use connect and per-read timeouts without a total response deadline.
• Raise generated/example stream_idle_timeout_ms from 120 seconds to 360 seconds.
• Add regression tests for idle SSE keep-alives, generated provider timeout settings, and streams that outlive the normal total timeout.

Validation

• cargo test --all-targets --quiet
• cargo test --lib --quiet
• cargo check --all-targets --quiet

All passed locally.