Reintroduce progressive HTTP read timeout support

[BGQ99]

What problem does this PR solve?

Issue Number: Related to apache#3133, follow-up to apache#3409 and apache#3453

Problem Summary:

Progressive HTTP response reads currently have no independent idle timeout after ReadProgressiveAttachmentBy() is called. A reader may therefore wait indefinitely when the peer stops sending body data while keeping the connection open.

PR apache#3409 introduced progressive-read timeout support, but post-merge CI exposed a lifecycle and synchronization problem around the timer callback and reader completion path. In particular, the timer callback could race with reader completion/error delivery, which made HttpTest.progressive_read_timeout_preserves_reader_error unstable and could replace the reader's original error with the timeout status.

This PR restores the feature with explicit synchronization and independent timer-state ownership.

What is changed and the side effects?

Changed:

• Add Controller::set_progressive_read_timeout_ms() for progressively read HTTP/1.x response bodies.
• Treat the configured value as the maximum idle interval between successfully consumed body parts.
• Start the watchdog when ReadProgressiveAttachmentBy() is called.
• Refresh the idle deadline after every successful reader callback.
• Do not count time spent inside the user reader callback as idle time.
• Preserve an error returned by the user reader instead of replacing it with a later timeout.
• Keep timer state independently alive with shared ownership, so the timer callback does not access a destroyed reader wrapper.
• Protect reader_failed, timeout_triggered, end_delivered, timer ownership and user-callback state with the same mutex.
• Cancel or detach the watchdog before delivering OnEndOfMessage().
• Close the HTTP/1.x socket through Socket::SetFailed() instead of destroying the parsing context from the timer thread.
• Use the versioned SocketId carried by HttpContext to avoid operating on a recycled socket.
• Return EPROGREADTIMEOUT when an idle timeout occurs.
• Reject progressive-read timeout for HTTP/2 with ENOTSUP.
• Update the HTTP client/server examples.
• Add focused unit tests covering:
  • active progressive reads;
  • idle timeout after receiving data;
  • timeout before the first body part;
  • slow user callbacks;
  • preservation of reader errors;
  • rejection for HTTP/2.

Side effects:

• Performance effects:

  • The feature is disabled by default.
  • When enabled, one watchdog timer and a small shared state object are created for the progressive reader.
  • Reader/timer state transitions use a mutex.
  • Existing calls that do not configure progressive_read_timeout_ms keep the original behavior.

• Breaking backward compatibility:

  • No source-level behavior change when the option is unset.
  • The new fields change the layout of brpc::Controller, so applications should be rebuilt when upgrading instead of mixing binaries built against different brpc versions.

Verification

Verified on commit fc865456fcc21f2e8325e3e9b0c27f7d6fce7b21.

Local focused stress tests:

• ASan: 200/200 rounds passed.
• 6 tests per round, 1,200 total test invocations.
• 0 test failures.
• 0 AddressSanitizer errors.
• Bazel 8.3.1: 20/20 runs passed.
• 6 tests per run, 120 total test invocations.

GitHub Actions:

• License Check: passed.
• Build on macOS: passed.
• Build and Test on Linux: passed.
• The latest complete Linux attempt passed all 11 jobs, including ASan and Bazel jobs.

Detailed sanitized logs, reproduction scripts and SHA256 checksums will be attached in a PR comment.

---

Check List:

• The changes are compilable.
• Focused unit tests were added.
• ASan stress testing passed.
• Bazel repeated testing passed.
• GitHub Actions passed.
• The changes follow the Contributor Covenant Code of Conduct.

[BGQ99]

Additional validation evidence for commit fc865456fcc21f2e8325e3e9b0c27f7d6fce7b21.

Local stress testing

• ASan: 200/200 rounds passed.
• 6 focused progressive-read timeout tests per round.
• 1,200 total test invocations.
• 0 failures.
• 0 sanitizer errors.
• Bazel 8.3.1: 20/20 runs passed.
• 120 total focused test invocations.

The attached archive contains:

• sanitized ASan log;
• sanitized Bazel log;
• ASan reproduction script;
• Bazel reproduction script;
• SHA256 checksums for every included file.

Archive:

progressive-read-timeout-test-evidence-fc865456.zip

Archive SHA256:

3641592B51D1BCE19F5BA1796FB2F91C27E5980D94CC9E040847D16AF2CFBAA4

GitHub Actions reruns

Five of six complete Linux workflow attempts passed.

Attempt #5 failed only in the existing
clang-unittest-bazel-with-babylon-and-new-pb job, in
RdmaTest.v3_server_rejects_oversized_pb_size.

The other 10 Linux jobs in that attempt passed. This PR does not modify RDMA
code or that test. The subsequent complete attempt #6 passed all 11 Linux jobs,
including the previously failing job.

• Attempt #5
• Attempt #6 — all jobs passed

Validation Screenshots

ASan stress test: 200/200 rounds passed

Bazel 8.3.1 stress test: 20/20 runs passed

GitHub Actions attempt #6: all Linux jobs passed