collision_gjk: remove norm2 array and compute inline

[thowell]

improve EPA memory utilization by computing a dot product of 3-vectors inline instead of storing the result

aloha_pot

mjwarp-testspeed benchmark/aloha_pot/scene.xml --nconmax=24 --njmax=128

this pr

Summary for 8192 parallel rollouts

Total JIT time: 0.69 s
Total simulation time: 4.17 s
Total steps per second: 1,962,558
Total realtime factor: 3,925.12 x
Total time per step: 509.54 ns
Total converged worlds: 8192 / 8192

main (24e07de)

Summary for 8192 parallel rollouts

Total JIT time: 0.66 s
Total simulation time: 4.18 s
Total steps per second: 1,959,876
Total realtime factor: 3,919.75 x
Total time per step: 510.24 ns
Total converged worlds: 8192 / 8192

throughput remains about the same