stderr of server threads swallowed during integration tests

[cole-miller]

Recently I haven't been able to get any tracing or other stderr output from the dqlite server thread(s) when running our integration tests, even when --show-stderr is passed. (stderr of the "client" thread that runs the test harness shows up as expected.) Apparently @MathieuBordere has run into this too. We should fix it!

[cole-miller]

This might have something to do with integration tests inadvertently linking to an installed libdqlite instead of the just-built one, since I was able to "fix" it by blowing away the lib directory where those artifacts were installed. A good sanity check is to replace the definition of tracef by

#define tracef(...) do { \
		if (UNLIKELY(_dqliteTracingEnabled)) { \
			__builtin_trap(); \
		} \
	} while (0)

[cole-miller]

I wonder if the _dqliteTracingEnabled static is getting duplicated somehow.