rpc: de-flake Server.OverloadTest{,NoCreateSpecialMessage}

[chen3feng]

Problem

Server.OverloadTestNoCreateSpecialMessage (and OverloadTest) flake on busy CI runners — the Bazel lane failed consistently with server_test.cc:261: ASSERT_EQ(100, succeeded) → Which is: 300 / 400, followed by a SIGSEGV in the next test.

The handler slept a fixed 2s, and the test fired N (1000/10000) calls at a 100-slot cap expecting exactly 100 to succeed. On a slow runner, submitting all N takes longer than 2s, so the first sleepy handlers finish and free slots mid-submit → more than the cap succeed. And because it's a fatal ASSERT_, the failure skipped gate->Stop()/server.Stop(), leaving in-flight calls → the SIGSEGV in the following test.

Fix — make it deterministic

SleepyEchoService now holds its concurrency slot until released (bumping an admitted counter on entry). Each test:

1. fires exactly the cap (100) with a long deadline (the held calls),
2. waits until the cap is full (admitted == 100),
3. floods the remaining N−100 with a short deadline — all dropped, since the cap is held,
4. waits for those to drain, then releases the held calls.

So precisely the cap's worth succeed, no extra slips into a freed slot, and nothing times out — independent of runner speed. Also switched ASSERT_EQ → EXPECT_EQ so a failure can't skip teardown.

Verification

• 10/10 passes under heavy CPU load (half the cores saturated) — was consistently failing on the busy CI runner
• ~1.1s per test (down from ~3.1s)
• all 10 server_test cases pass; the downstream SIGSEGV is gone