Remove CUDA syncs from circular buffer

[rdeits-bd]

I was doing some CUDA profiling of our RL environments and found some surprising CUDA syncs associated with actuator delays. I tracked them down to these two places in circular_buffer.py, in which the attempt to skip an unnecessary op actually slowed things down by forcing a CUDA sync. Note that I'm still pretty new to this; I'm just following https://docs.nvidia.com/dl-cuda-graph/torch-cuda-graph/sync-free-code.html

In our internal example, I got a ~3% speedup from this change, though I suspect the effect size will vary a lot based on whether the rest of the MDP is sync-free.

@kevinzakka I'll try to run your benchmarks from #1020

[rdeits-bd]

Ok, got some results for the Go1 velocity task. Here's what I get on #1020 for the Go1 task (with no actuator delay, as is the default):

==========================================================================
Summary (all values in env steps per second):
  Physics SPS: sim.step() only (×decimation per env step)
  Env SPS: full env.step() including managers
  Overhead: time spent on non-physics work (observations, rewards, etc.)
==========================================================================
Task                                 Dec  Physics SPS      Env SPS Overhead
--------------------------------------------------------------------------
Velocity-Flat-Go1                      4      641,353      127,292    80.2%

If I set delay_min_lag = delay_max_lag = 1 for the Go1 articulation, then I get:

==========================================================================
Summary (all values in env steps per second):
  Physics SPS: sim.step() only (×decimation per env step)
  Env SPS: full env.step() including managers
  Overhead: time spent on non-physics work (observations, rewards, etc.)
==========================================================================
Task                                 Dec  Physics SPS      Env SPS Overhead
--------------------------------------------------------------------------
Velocity-Flat-Go1                      4      644,432      112,096    82.6%

a pretty significant drop in environment SPS.

With this PR on top of yours, I get:

==========================================================================
Summary (all values in env steps per second):
  Physics SPS: sim.step() only (×decimation per env step)
  Env SPS: full env.step() including managers
  Overhead: time spent on non-physics work (observations, rewards, etc.)
==========================================================================
Task                                 Dec  Physics SPS      Env SPS Overhead
--------------------------------------------------------------------------
Velocity-Flat-Go1                      4      658,344      118,552    82.0%

i.e. recovering about half of the performance loss.