panfrost: Multi-threading and fragment_ops fixes

This gets dEQP-EGL.* passing. We're not race-free after this, unfortunately, but the most egregious races are fixed (i.e. everything the CTS hits.)

Still some races with panfrost_resource which should be debugged.

Thank the stars for helgrind.

This also includes fixes for dithering and multisampling. Throwing this branch through the wringer tonight.