faster + nicer visualization

[latentCall145]

Inspired by @gujialiang123's new visualizer but based off main and with some (opinionated) changes:

• Use InstancedMesh to massively speed things up (previously ran < 60 FPS, now runs 60 FPS with my CPU < 20%)
• blanked out Program ID slider based on the kernel launch grid (e.g. for grid (1, 2, 4), there's no slider needed for the X dimension as there's only one program launched on that dim)
• fixed some issues
  • hovering over an item gave wrong value (due to some dimension permutation issue)
  • clicking on multiple records used to highlight all of them instead of just the actively viewed record
• removed stuff I didn't find helpful
  • "Drag Cubes" button removed
  • "Show Code" button removed (code now always shown)
  • "Slice Tensor" removed
  • load/store/dot animations (e.g. yellow cubes in "before" preview)
  • Any features that aren't accessible through the main visualizer (e.g. flip animation, calibrate, etc.)

preview

before (when rotating view, frames render in ~170 ms on my laptop)

examples/3dims.py

examples/matmul_demo.py

after (when rotating view, frames render in ~2-3 ms, changing program ID takes ~20-30 ms)

examples/3dims.py

examples/matmul_demo.py

note: I'm currently working on a new visualizer (and tracer) to address some usability issues in the current one (e.g. I'd like to click on a line of source code and visualize the correct record similar to nsight compute, record repeated calls to the same line of code separately vs. call to a different line of code, allow others to submit a trace file in the web demo rather than running the tracer right before the visualizer, etc.), this PR is just something in case I don't have anything for the new visualizer + tracer