MVKCmdDrawIndirect: Fix indirect index for triangle fans.

[cdavis5e]

Analogously to #2419, this fixes generation to add vertexStart to the generated indices, and start from zero in the array.

[dboyan]

I'm not 100% sure yet if this is correct. The case with drawCount>1 might be interesting, but currently I haven't tested it. I tested drawCount=1 case with my own minimal program (written in gles and hooked to angle's vulkan backend because I'm not a vulkan expert), via glDrawArraysIndirect. It rendered correctly with or without the change (I guess separate calls don't suffer from buffer overwriting problem here). But I couldn't verify the more generic case as I didn't have the extensionGL_EXT_multi_draw_indirect with angle. I'll see if I can obtain a vulkan version of test program.

[dboyan]

I was able to get a working test program and now I can confirm the change isn't in the right direction. Actually the original code makes sense somewhat.

Let's consider the situation where we have 16 vertices, and we call vkCmdDrawIndirect with drawCount=2, using triangle fan primitive topology. The indirect buffer contains

Item 0: { vertexCount=4, firstVertex=0 }
Item 1: { vertexCount=4, firstVertex=8 }

Notice that the compute kernel cmdDrawIndirectPopulateIndexes handles both draws with 2 threads, and the resulting index are populated into a single buffer idxBuff. With the original logic, idxBuff will contain

[0, 1, 2, 3, X, X, X, X, 8, 9, 10, 11, X, X ...]    (X means uninitialized element)

This is the intended behavior because indices from multiple draws are put "in place" and will not interfere with each other. They will be further processed to generate indices for triangles according to the destBuff output of the current stage. It seems a waste of space but it's logically correct.

On the other hand, after the change, all the indices from different draw all start from the beginning of idxBuff, and that creates trouble. My #2419 was correct because vkCmdDraw can only issue a single draw, so one doesn't need to worry about the issue above.

In fact, even with the original code, the rendering is not 100% correct. The triangle fan translation seems a little off. I'll open a issue when I get time.

[cdavis5e]

Yeah, I realized a bit after the fact that it was already adding the base before my change. Thanks for confirming.