Fix GIL release and acquire when embedding the interpreter

Fixes a race condition when multiple threads try to acquire the GIL
before `detail::internals` have been initialized. `gil_scoped_release`
is now tasked with initializing `internals` (guaranteed single-threaded)
to ensure the safety of subsequent `acquire` calls from multiple threads.

tests/test_embed/test_interpreter.cpp

@@ -1,6 +1,8 @@
 #include <pybind11/embed.h>
 #include <catch.hpp>
 
+#include <thread>
+
 namespace py = pybind11;
 using namespace py::literals;
 
@@ -185,3 +187,32 @@ TEST_CASE("Execution frame") {
     py::exec("var = dict(number=42)");
     REQUIRE(py::globals()["var"]["number"].cast<int>() == 42);
 }
+
+TEST_CASE("Threads") {
+    // Restart interpreter to ensure threads are not initialized
+    py::finalize_interpreter();
+    py::initialize_interpreter();
+    REQUIRE_FALSE(has_pybind11_internals_static());
+
+    constexpr auto num_threads = 10;
+    auto locals = py::dict("count"_a=0);
+
+    {
+        py::gil_scoped_release gil_release{};
+        REQUIRE(has_pybind11_internals_static());
+
+        auto threads = std::vector<std::thread>();
+        for (auto i = 0; i < num_threads; ++i) {
+            threads.emplace_back([&]() {
+                py::gil_scoped_acquire gil{};
+                py::exec("count += 1", py::globals(), locals);
+            });
+        }
+
+        for (auto &thread : threads) {
+            thread.join();
+        }
+    }
+
+    REQUIRE(locals["count"].cast<int>() == num_threads);
+}