re_renderer: implement depth cloud renderer

[teh-cmc]

This implements a new Renderer within re_renderer that takes a depth texture and a set of intrinsics as input, and draws a point cloud as output.
It also adds a new example which compares doing backprojection into a point cloud manually vs. using the new depth-cloud renderer.

From a rendering standpoint, this work very similarly to the existing point renderer, except that the vertex shader backprojects a depth texture rather than reading from a position/color texture.
The depth texture is uploaded as-in and just-in-time, no caching involved.
Alpha and shading are both turned off.

This PR does not include a bunch of things on purpose as it's already way too big as-is (though most of it is boilerplate).
All of these will come in (hopefully much smaller) follow up PRs:

• integration with main Rerun app
• support for picking which primitive to render (image vs. point cloud)
• support for (gpu) color maps
• support for albedo textures

Picking is ignored altogether, until we have a generalized solution for that.

---

23-02-27_11.05.43.patched.mp4

23-03-01_22.21.49.patched.mp4

[teh-cmc]

This doesn't show anything on the web @Wumpf (see screenshot) which I'm sure is yet another case of WebGL alignment shenanigans... If you've got an idea I'd love to hear it 👀

[Wumpf]

alignment issues are very unlikely at this point for web-only problems. Also when I had issues with that on the web it was a crash. Anything useful in the log? Will have a closer look myself later!

[Wumpf]

this checks in very large binary files, we need to generate some synthetic data instead

[emilk]

this checks in very large binary files, we need to generate some synthetic data instead

Yeah, let's avoid checking in binary blobs 😬

• Large repo size #1403

[teh-cmc]

I'll make a procedural one.

[Wumpf]

Anything useful in the log?

No, because it doesn't yet have #1385 which reduces a lot of log spam 😅

[Wumpf]

Nice start!

I'm a bit worried about the amount of boiler plate needed in some places, but can (should) take that separately.

Main issue is about Mat3 as already pointed out on Slack.
Other issues are mostly about naming and suggestion for code-reuse.

Best candidate for WebGL issue is the use of D16 sampling which I found somewhat exotic (but after getting past my first "shock" is totally watertight with the spec and the correct way to go about it as I believe now!).

[teh-cmc]

See also gfx-rs/wgpu#3537

[teh-cmc]

Should be good to go!

[Wumpf]

let's do this! :)