Fix benchmark

[ThierryBerger]

Related to #553

• Adding a simple Transform wasn't enough to place correctly the cubes, which made the scene having way too many exploding collisions, which was the main source of bad performance detected in Added initial benchmarks #551.
  • Now the performance is more comparable to rapier testbed benchmark. There is a overhead but not a horrible one:
    • bevy takes ~5 seconds to boot up, not sure exactly why, we might be able to strip a few plugins (bevy_render, window... ) but I didn´t succeed, async-collider was complaining, even though I believe I deactivated it 🤔.
    • on 1000 iteration on 40 pyramids with 20 height, bevy takes 26 seconds once booted up where rapier takes 20 seconds. So there is still room for improvement.
• I also added a feature "visual" to more easily see what the benchmark is doing. it's mostly meant to be used while working on the benchmarks, enabling only 1. We cannot run them sequentially because bevy_winit doesn´t support window recreaction.

[ThierryBerger]

add explaination to readme

[sebcrozet]

Just need the README update for the visual feature, but looking good overall. Thanks!

[sebcrozet]

This reminds me that I’d like to revisit that profiler thing in rapier to switch so something less manual. Like switching the tracy or something. (This is unrelated to this PR though.)

[ThierryBerger]

Running cargo run --release --bin bench from the root brings all dependencies, but running this from bevy_rapier_benches3d brings only the needed ones, this may be worth a note (and I'd like to understand why.)