Switch back to a more compact line attributes layout

[mourner]

Switch back from 12 to 8 bytes per vertex for line layers. See #8304.

I tried to decode them in a potentially safer way this time (avoiding mod and / in shader math) to see if it fixes the Broadcom GPU bug — we need to verify this before merging we did verify it works properly on Broadcom VideoCore IV GPU, and faster than master 🎉. If not, this PR will be a good starting point for exploring other packing strategies (e.g. packing in most significant bit).

Launch Checklist

• briefly describe the changes in this PR
• write tests for all new functionality covered by render tests
• document any changes to public APIs no changes
• post benchmark scores ran Layout/Paint/LayerLine, they don't seem to be affected since this is a GPU thing
• manually test the debug page
• tagged @mapbox/gl-native if this PR includes shader changes or needs a native port
• test on the affected Broadcom GPU device
• test GL Native counterpart on a Broadcom GPU device [core] Switch back to a more compact line attributes layout mapbox-gl-native#14851

[ansis]

The testing so far looks promising but I think we should try to test this in -native with the buggy device before merging.

Do you think using the most significant bit instead of the least significant bit could make it even more robust?

[ansis]

Would a new buffer size benchmark be useful to track these kinds of changes?

[mourner]

The testing so far looks promising but I think we should try to test this in -native with the buggy device before merging.

Yep, already submitted a PR mapbox/mapbox-gl-native#14851 — will wait for confirmation of that one.

Do you think using the most significant bit instead of the least significant bit could make it even more robust?

I don't know why it wasn't robust in the first place so not sure whether it will make a difference, but if the current approach is confirmed to work with those GPUs, we shouldn't have any problem on newer ones.

Would a new buffer size benchmark be useful to track these kinds of changes?

Yes, I have it locally and want to PR, but it doesn't fit the benchmark format because it's about qualitative values, not timing. Might be worth pushing just as a debug page at first.