Integration test downstream client crashes when a long send is interrupted with an early close

Discussed this crash with @htuch before making public

Title: Integration test downstream client crashes when a long send is interrupted with an early close

Description: I encountered the following crash while fuzzing the ext_proc filter.

If the downstream client is making a large POST request in chunks and Envoy responds before the downstream client is finished sending, a seg fault occurs.

Adding the following test case to ext_proc's streaming_integration_test can recreate this scenario:

// Perform a long chunked send while the external process immediately responds to the headers
TEST_P(StreamingIntegrationTest, LongSendCrash) {
  const uint32_t num_chunks = 150;
  const uint32_t chunk_size = 100;

  test_processor_.start(
      [](grpc::ServerReaderWriter<ProcessingResponse, ProcessingRequest>* stream) {
        ProcessingRequest header_req;
        ASSERT_TRUE(stream->Read(&header_req));
        ASSERT_TRUE(header_req.has_request_headers());

        ProcessingResponse header_resp;
        auto* immediate = header_resp.mutable_immediate_response();
        immediate->mutable_status()->set_code(envoy::type::v3::StatusCode::Unauthorized);
        immediate->set_body("{\"reason\": \"Not authorized\"}");
        immediate->set_details("Failed because you are not authorized");
  	stream->Write(header_resp);
      });

  initializeConfig();
  HttpIntegrationTest::initialize();
  sendPostRequest(num_chunks, chunk_size, absl::nullopt);

  // TODO skipping checks of response
}

Additional Information: I'm not sure if this is specific to ext_proc.

The test is fairly flaky, you can increase the num_chunks value if the issue isn't being recreated on your machine.

Further, I could only recreate the crash after executing all tests in the test suite.

Logs: I have provided truncated.log which shows the crash, but I removed the output from the other tests in the suite.