Remove WebIO and build directly on Mux.jl and HTTP.jl

[rdeits]

I've been considering this for a whlie, and I think it's time to do it. MeshCat doesn't actually gain much from being built on top of WebIO, and there have been a fair number of drawbacks, mostly just the churn of having a pretty complicated dependency in the stack. This PR removes the WebIO layer and re-implements the meshcat protocol directly with Mux.jl, HTTP.jl, and WebSockets.jl. Those are the same packages that power WebIO under the hood, so nothing is changing except the way we interact with those tools. We even get to maintain compatibility with WebIO (by implementing WebIO.render with Requires.jl), so integration with other WebIO tools should keep working.

What works:

• Creating and opening visualizers in the browser
• IJulia cell integration
• Blink window integration
• Juno plot pane integration
• VSCode plot pane integration
• ElectronDisplay.jl integration
• animations
• WebIO.render() integration (now handled via a plain iframe and lazily loaded via Requires.jl)

What doesn't work:

• Serving visualizers from remote jupyter servers (e.g. NextJournal). This was the primary motivation for the switch to WebIO, but I've never actually used it successfully. I don't think this hypothetical feature is worth all the complexity it requires. And in fact, I have an idea for how we can make this work anyway: rather than creating a live connection from the remote server to the client, we can create a new function which sends a fully static rendering of the entire scene, including any animations. That won't deliver the full MeshCat experience, but it should be pretty stable and cover most cases.
• The whole AbstractControls infrastructure. This never worked anyway, so I'm removing it.

Pros:

• Better performance: WebIO required us to re-encode our data as JSON, even though we were already streaming it as msgpack. Dropping WebIO means we can use msgpack directly, which makes a huge difference for big meshes and pointclouds.
  • To show this, I've increased the number of points in the demo notebook pointcloud from 100k to 1 million
• Jupyter integration just works (locally). No need to install the WebIO jupyter extension, which has been tricky to get right.
• Jupyterlab just works (fixes Meshcat works in jupyter notebook but does not in jupyterlab #83 )
• We're now rendering the same index.html that the python bindings use. That makes life easier on me (and fixes Might be nice to have a global viewer handle for debugging #33 )
• More control over logging verbosity (fixes Closing visualizer tab results in websocket connection error #108 )
• Less sketchy async behavior (fixes with_logger(NullLogger()) breaks MeshCat #68 )
• Draw commands are now synchronous by default (and can be @asynced to be async) (fixes More synchronous operations #60 )

Cons:

• Slightly more boilerplate for IJulia and Blink? Edit: not really
• Probably won't work in Juno without some additional integration. Edit: this works now
• Might require some adjustment to work with the RigidBodySim GUI stuff

[codecov-io]

Codecov Report

Merging #141 into master will increase coverage by 2.12%.
The diff coverage is 55.17%.

@@            Coverage Diff             @@
##           master     #141      +/-   ##
==========================================
+ Coverage   79.27%   81.39%   +2.12%     
==========================================
  Files          15       18       +3     
  Lines         468      473       +5     
==========================================
+ Hits          371      385      +14     
+ Misses         97       88       -9

Impacted Files	Coverage Δ
src/lowering.jl	98.48% <ø> (+8.2%)	⬆️
src/commands.jl	83.33% <ø> (+11.9%)	⬆️
src/assets.jl	0% <0%> (ø)
src/servers.jl	4% <0%> (-0.55%)	⬇️
src/integrations.jl	100% <100%> (ø)
src/MeshCat.jl	80% <100%> (+46.66%)	⬆️
src/render.jl	33.33% <33.33%> (ø)
src/visualizer.jl	91.42% <90.69%> (+15.37%)	⬆️
... and 1 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update b4d0be1...d5d75bf. Read the comment docs.

[rdeits]

@travigd I wanted to let you know about this. Basically, I reimplemented the core MeshCat interface directly in Mux/HTTP/WebSockets, rather than using WebIO to handle communications and asset serving. This helped with comms performance (by skipping the JSON encode) and overall reduced the complexity of the system a bit. Integration with other WebIO stuff should keep working, since MeshCat now defines WebIO.render via Requires.jl. I suspect this will be a better way for these tools to coexist, and it means that MeshCat should be much less sensitive to internal changes in WebIO. Good news all around 🙂

[rdeits]

@tkoolen this might slightly break the RigidBodySim GUI controls, but I think it will be easy to fix. It also might just work as-is, since I'm still defining WebIO.render. Anyway, I'll make sure this is a new minor version, so we can update RigidBodySim later.