Fix GC profiling timing

We were recording GC profile timestamps inside the VM postponed job that
flushes temporary information.  The postponed job could be run much
later than the sample was actually taken, so this commit records a
timestamp in the signal handler when there is a GC event.

It only records one time stamp, so if there are multiple GC events in a
row, we have to assume that the total duration spent in GC starts from
the first time stamp and extends until the timestamp of the next non-GC
sample.  In other words, we don't record a correct timestamp for each GC
sample, only the first one.

ext/stackprof/stackprof.c

@@ -120,6 +120,8 @@ static struct {
     size_t unrecorded_gc_sweeping_samples;
     st_table *frames;
 
+    timestamp_t gc_start_timestamp;
+
     VALUE fake_frame_names[TOTAL_FAKE_FRAMES];
     VALUE empty_string;
 
@@ -646,15 +648,15 @@ stackprof_buffer_sample(void)
     _stackprof.buffer_time.delta_usec = timestamp_delta;
 }
 
+// Postponed job
 void
 stackprof_record_gc_samples(void)
 {
     int64_t delta_to_first_unrecorded_gc_sample = 0;
     uint64_t start_timestamp = 0;
     size_t i;
     if (_stackprof.raw) {
-	struct timestamp_t t;
-	capture_timestamp(&t);
+	struct timestamp_t t = _stackprof.gc_start_timestamp;
 	start_timestamp = timestamp_usec(&t);
 
 	// We don't know when the GC samples were actually marked, so let's
@@ -776,6 +778,10 @@ stackprof_signal_handler(int sig, siginfo_t *sinfo, void *ucontext)
 	} else if (mode == sym_sweeping) {
 	    _stackprof.unrecorded_gc_sweeping_samples++;
 	}
+	if(!_stackprof.unrecorded_gc_samples) {
+	    // record start
+	    capture_timestamp(&_stackprof.gc_start_timestamp);
+	}
 	_stackprof.unrecorded_gc_samples++;
 	rb_postponed_job_register_one(0, stackprof_job_record_gc, (void*)0);
     } else {