VertexLoader: do not prepare for vertices if we need to skip them

[delroth]

Fixes a recent regression from #716. Not a big fan of doing the DataSkip in VertexLoaderManager, but tbh it isn't much worse than doing it in VertexLoader. It also removes the dependency of VertexLoader on bpmem, which makes me happy.

[delroth]

Tested by @JMC47 to fix the regression, btw.

[lioncash]

LGTM

[degasus]

@lioncash I always fail here at the naming. Does "all" means all are passed or all are skipped? But in the end, the one case we filter here is the only uncommon one, so you'll notice such a issue very soon.

btw, as this case is quite uncommon, it's fine imo to move this skip where ever we want, eg VertexManager::Flush (which is already based on BP)

[degasus]

btw, I think the reason for this "hack" here is that opengl (maybe also d3d) doesn't support this culling mode (-> render nothing): http://www.opengl.org/wiki/Face_Culling

[lioncash]

@degasus I don't know, I'm just going off of the enum here

[neobrain]

@degasus "cull all" means all of the triangles are culled away (i.e. none of them are used for rendering). Indeed, this is required because OpenGL doesn't support it.

It might be useful for reference to know that this mode is used to set reference planes for zfreeze without actually rendering things on screen. Note that due to this it would be a severe problem for the software renderer if it had used the code including the workaround without further ado.