load stats: fix integration test flake

[mattklein123]

Waiting on a load stats response can race with resetting
the counters when initializing a watch. Moving the counter
increment prevents the race.

Fixes #11784

Risk Level: Low
Testing: Existing tests, 1000 runs of the integration test
Docs Changes: N/A
Release Notes: N/A

[htuch]

The change LGTM (do no harm), but I don't fully grok why this fixes the race. Can you elaborate a little more for posterity?

[mattklein123]

There is a race condition here:

envoy/test/integration/load_stats_integration_test.cc

Lines 603 to 605 in da58cfc

	requestLoadStatsResponse({"cluster_0"});
	initiateClientConnection();
	ASSERT_TRUE(waitForLoadStatsRequest({localityStats("winter", 0, 0, 1, 1)}));

requestLoadStatsResponse returns when "load_reporter.requests" is incremented. After it's incremented, the host stats are cleared (in particular rq_total). Thus, it's possible to miss the rq_total increment if the client connection happens after the stat increment but before the clear.

Sorry I should have made this more clear. This took me a while to figure out.

[htuch]

Makes sense, thanks!