gh-150191: Avoid data race in test_sni_callback_race

[colesbury]

Use a single handshake thread to avoid OpenSSL-internal data races on shared SSL_CTX state, while keeping multiple togglers to exercise the sni_callback setter race from gh-149816.

• Issue: Data race in test_ssl.test_sni_callback_race #150191

[colesbury]

A few other related changes:

• Run the test in the GIL-enabled build as well. "test is only useful if the GIL is disabled" is wrong. The data race reported exists in the GIL-enabled build as well. In general, we should write these tests so that they run in both the GIL-enabled and free-threaded build.
• Use threading_helper.run_concurrently, which sets up the needed exception tracking and uses a barrier, which helps catch data races by making it more likely that threads run at the same time.
• Run the threads for 0.1 seconds. Previously "done.set()" was called immediately so that the threads barely had time to run, if at all. (Alternatively, we could run them for a fixed number of iterations)

[kiri11]

@colesbury Thanks for the test improvements!

• Run the threads for 0.1 seconds. Previously "done.set()" was called immediately so that the threads barely had time to run, if at all. (Alternatively, we could run them for a fixed number of iterations)

Running it for at least 0.1s does dramatically increase the consistency, although I've been told we shouldn't add static delays to the tests, even as small as 0.1s.