I'm not sure why values_capacity_ is different kept separately
from the buffer, but there is check which does not reserve
capacity again values_capacity_ is already the needed size.
When ReleaseValues is called, we allocate a brand new empty buffer.

I'm not really sure why this hasn't caused users more issues (maybe
increasingly large row groups or some other phenonemon). This bug
also highlight that our C++ tests have very limited coverage on batched
reads.  To fix this I added an batch read for every round trip test
to confirm it yields the same values.