Fix CPU profiler latency

The current profile must be stopped AFTER the next begins. This is because V8 will
tear down the symbolizer thread when the active profile count reaches zero. This
means the next profile started will spin up a new symbolizer thread. When this
thread starts it streams in all the JIT location information which could block the
JavaScript thread for a significant length of time.

packages/dd-trace/src/profiling/profilers/cpu.js

@@ -25,9 +25,7 @@ class NativeCpuProfiler {
 
   profile () {
     if (!this._stop) return
-    const profile = this._stop()
-    this._record()
-    return profile
+    return this._stop(true)
   }
 
   encode (profile) {

packages/dd-trace/test/profiling/profilers/cpu.spec.js

@@ -74,7 +74,7 @@ describe('profilers/native/cpu', () => {
     expect(profile).to.equal('profile')
 
     sinon.assert.calledOnce(stop)
-    sinon.assert.calledTwice(pprof.time.start)
+    sinon.assert.calledOnce(pprof.time.start)
   })
 
   it('should encode profiles from the pprof time profiler', () => {