When the external buffer of a surface is dirty, and we're trying to
lock the internal buffer (or vice-versa), an 'update' needs to happen.
This resulted in us locking the internal buffer for writing, thereby
marking it as dirty. This triggered an assert which checks that both
buffers can't be dirty at the same time. Also, in release mode this
could result in redundant updates when the external buffer is locked
again.

We should mark the destination buffer dirty when we're about to make
sure it's up to date, so a new lock type was added which is equivalent
to a write lock but without setting the dirty flag.

Also, we were allocating memory for a 0x0 stencil buffer for each
render target, when there's no stencil component. So return nullptr
when the format is NULL.

Change-Id: Ie7b5528e3eedc3c3efdf8461047e6284b7bdfc84
Reviewed-on: https://swiftshader-review.googlesource.com/16828Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>