Expand RS vector3 types to vector4.

BUG=5609007

The underlying LLVM implementation for vector3 types does this implicitly. If
RS does not adjust its implementation, we will always be misaligned for any
subsequent data after a vector3 type. We previously inserted padding into the
reflected layers from llvm-rs-cc (hence the skip padding part of this change).
We can safely ignore the padding now that the Java/native code is updated to
use the expanded size. The compiler will also need modification to ensure that
we don't mistakenly skip over any end-of-struct padding.

Fixing the 3 component vector padding problem.

Change-Id: If68af42287deb8f4b28addcd19a9fa314656be44
diff --git a/libs/rs/rsComponent.cpp b/libs/rs/rsComponent.cpp
index 7d9cf0b..21b98f6 100644
--- a/libs/rs/rsComponent.cpp
+++ b/libs/rs/rsComponent.cpp
@@ -169,7 +169,8 @@
         break;
     }
 
-    mBits = mTypeBits * mVectorSize;
+    mBitsUnpadded = mTypeBits * mVectorSize;
+    mBits = mTypeBits * rsHigherPow2(mVectorSize);
 }
 
 bool Component::isReference() const {