some rayTest improvements in btDbvt::rayTestInternal, it avoids/reduces memory allocations during stack allocation (by sharing a persistent m_stack)

and rayTestInternal re-uses precomputed invRayDirection/signs.
also did some performance comparison with different ray-AABB test, from 
http://jgt.akpeters.com/papers/EisemannEtAl07/

In short: it is faster, but it is not clear how to cull ray segments using ray slopes: when rays starts inside the AABB, we get a negative t value, but negatives also get false t-values...

src/BulletCollision/BroadphaseCollision/btBroadphaseInterface.h

@@ -21,10 +21,23 @@ subject to the following restrictions:
 struct btDispatcherInfo;
 class btDispatcher;
 #include "btBroadphaseProxy.h"
+#include "LinearMath/btAabbUtil2.h"//for fast ray-slope algorithm
+
 class btOverlappingPairCache;
 
+
+
 struct	btBroadphaseRayCallback
 {
+	///added some cached data to accelerate ray-AABB tests
+	///m_ray is used to accerate ray-AABB tests, see btDbvt.h, when TEST_RAY_SLOPES is enabled in LinearMath/btAabbUtil2.h
+	btRaySlope	m_ray;
+
+	///otherwise this data is used to accelerate ray-AABB tests
+	btVector3		m_rayDirectionInverse;
+	unsigned int	m_signs[3];
+	btScalar		m_lambda_max;
+
 	virtual ~btBroadphaseRayCallback() {}
 	virtual bool	process(const btBroadphaseProxy* proxy) = 0;
 };