Fused scan kernel optimization

[jonahsamost]

We optimize this kernel with a sort of checkpointing (see CHECKPOINT_INTERVAL in the code). It reduces intermediate buffer writes/reads by 4x. The backward recomputes values between checkpoints instead of reading them. The forward only writes every CHECKPOINT_INTERVAL amount of times. Also use fast math intrinsics where we can. End up with 2.3x forward and 1.03x backward speedups (about ~1.3x overall).

Had to fix some build issues i had as well.

Tested on 5090

fused_scan (N=4194304, 512x64x128, combined=512x64x384)
  	forward             74.0 us  442.54 M elem/s
  	backward           106.1 us  308.80 M elem/s
  	forward (checkpointed)   31.6 us  1038.33 M elem/s
  	backward (checkpointed)  103.4 us  316.86 M elem/s
  checkpointed forward correctness: out=ok(3.29e-05) next_state=ok(3.58e-05)
  checkpointed backward correctness: grad_combined=ok(5.43e-05) grad_state=ok(4.77e-07)