PPO kernel optimization

[jonahsamost]

This kernel mainly attains a speed up from doing an online softmax in both the forward and backward passes and doing computation in 32 bit instead of 64 bit precision. It also avoids writing intermediate values to be used in the backward pass in favor of recomputation.

ppo_loss (NT=32768, 512x64, A=4)
  	forward (original)   17.3 us  1897.58 M elem/s
  	backward (original)   14.4 us  2279.77 M elem/s
  	forward (optimized)    3.3 us  9848.82 M elem/s
  	backward (optimized)    3.3 us  9787.83 M elem/s
  	forward (torch)     17.3 us  1890.14 M elem/s
  	backward (torch)   141.5 us  231.54 M elem/s
  	forward (cpp)      220.1 us  148.87 M elem/s
  	backward (cpp)    1084.5 us   30.21 M elem/s
  	forward (graph)     49.2 us  666.60 M elem/s