Replace OpenCV with GPU blob detection

[osnr]

Increases fps 3-4x (depending on dot size) and makes page load times way faster.
https://www.youtube.com/watch?v=7D3wY_wMGLo

Uses Laplacian of Gaussian at fixed scale to find dots; video data never needs to touch the CPU now.

todos

needs some extra docs on what sigma and dotThreshold mean

• sigma = expected dot radius in pixels
• dotThreshold = how strict to be in recognizing dots (gets rid of spurious dots)

should probably fix up outCanvas so it's not just hanging out there

would like to test on different cameras than 1080p, cc @ajbouh

other optimization ideas:

• async GL readPixels (readPixels is very expensive right now)
• simplify shader computation? one-pass Gaussian, one-pass Gaussian+Laplacian, difference of Gaussians, FFT and then do computation in frequency domain?
• shrink output canvas so we don't need to read as much from GPU?

[janpaul123]

Amazing. As discussed in person, going to bring this in a bit more incrementally, and next to the OpenCV implementation for now (since this is still less robust).

[janpaul123]

@osnr maybe the marker sizes from #53 are handy here?