Introduce ColorMode, StencilMode, DepthMode abstractions

[lbud]

Abstracts all depth-, color- and stencil-related state manipulation into "modes," so that each render function is responsible for setting its own modes (with context.set*Mode as an intermediate step, but eventually as part of program.draw). This eliminates all instances of render functions needing to reset assumed state after they're finished 🎉

I figured out all of the quirks I mentioned offline, so context.setStencilMode, context.setColorMode, and context.setDepthMode behave almost exactly the same as in native (the one slight difference being how we handle blendColor, which is set mostly as default-initialized in native, but is optional and often therefore not set here).

Benchmarks: https://bl.ocks.org/anonymous/raw/5a607b699c5f57ed705e4a7e2aa342a1/

The LayerBackground test does consistently look significantly slower (~1.89 ms, as opposed to previously ~0.59 ms) but I don't feel that's concerning given that most maps use one or very few background layer.

I don't feel comfortable merging this yet until I add a bunch of render tests that test various combinations of layers (which would have caught e.g. 5565087 — I'm going to add the render test for that to a new branch off of current master).

Refs #145.
Refs mapbox/mapbox-gl-native#10655.

[mourner]

Nit: all those modes look immutable (they're not modified once created), so should we make them constants instead of allocating new objects/arrays on every draw call?

[jfirebaugh]

There are two "masks" in the stencil state: the write mask (set by glStencilMask), and the function mask (set by the third parameter to glStencilFunc). This definition uses mask for both of them. The place where I'd expect this to create a problem is stencilModeForClipping, which should use 0 for the write mask and 0xFF for the function mask (it looks like native uses a dynamic value for the function mask -- not sure if this discrepancy matters).

In native, I added the SimpleTest and MaskedTest types to enforce that the function mask is omitted (because irrelevant) when the test function is NEVER or ALWAYS. The moral equivalent with flow types is something like:

type StencilTest =
   | { func: NEVER, mask: 0 }
   | { func: LESS, mask: number }
   | { func: EQUAL, mask: number }
   | ...
   | { func: ALWAYS, mask: 0 };

[jfirebaugh]

In native, a default-initialized color is rgba(0,0,0,0), which corresponds to the default value for glBlendColor. In GL JS you can use Color.transparent for the same purpose, and make blendColor non-nullable.

[jfirebaugh]

In native, RenderBackgroundLayer::render uses DepthMode::ReadOnly, which is equivalent to false. Is this a bug? (The ReadOnly and ReadWrite enum values are nice for readability -- can JS do something similar?)

[lbud]

Oh — actually both are wrong, I think. Since we draw opaque backgrounds in the opaque pass we should be writing to the depth mask only on that pass, as drawFill does. (I happened to try a test case with two opaque background layers, and they're rendered in the wrong order — only the bottom layer appears. Obviously this is an edge case that probably very few people would ever need to support, but 🤷‍♀️)

Re: ReadOnly and ReadWrite enums — 👍

[jfirebaugh]

This might be a good opportunity to fix mapbox/mapbox-gl-native#9713 (though it's ticketed in -native, we could reverse it here instead).

[lbud]

🤔 I tried doing this and must have been getting something wrong as most things no longer rendered correctly (if at all) — I'm going to merge this as-is and we can take on mapbox/mapbox-gl-native#9713 separately.

[jfirebaugh]

Any reason not to group this with the other set*Mode calls inside drawExtrusionTexture?