Optimized logcumsumexp

[jonahsamost]

Optimzed the logcumsumexp forward/backward by mainly moving things to float32 with no precision issues. In the forward I removed an unncessary branch. In the backward I added a fused multiply add.

logcumsumexp (N=4194304, 512x64x128)
  optimized correctness: forward=ok(2.38e-06), backward=ok(1.19e-06)
    forward            60.6 us  540.58 M elem/s
    backward          247.9 us  132.20 M elem/s
  forward (opt)       14.3 us  2286.18 M elem/s  4.23x speedup
  backward (opt)      10.2 us  3198.31 M elem/s  24.19x speedup
    forward (torch)    61.4 us  533.59 M elem/s
    backward (torch)  247.9 us  132.20 M elem/s
    forward (cpp)      28.7 us  1143.12 M elem/s
    backward (cpp)     89.1 us  367.96 M elem/s
    forward (graph)    25.6 us  1278.54 M elem/s
    ```