Use NVPerfWorks to measure FLOPs

[vchuravy]

Still need to verify that the data is accurate using ncu as a baseline,
and we need to decide if we want to average with something like OnlineStats.jl

This is measuring FLOP/s in "kernel time" not "host time".

Running locally with NZ=12 RESOLUTION=3

┌ Info: Duration
│   time = 1.44577531
│   gctime = 0.0
└   bytes = 13188120
┌ Info: Fraction active
└   fraction = 0.5
┌ Info: Kernel performance
│   time = 0.5812019466666665 s
│   D_FLOP = 1.00283101181e11 Instruction
│   F_FLOP = 7.03965312e9 Instruction
│   H_FLOP = 0.0 Instruction
│   D_FLOPs = 1.725443312021713e11 Instruction s^-1
│   F_FLOPs = 1.2112232521542831e10 Instruction s^-1
└   H_FLOPs = 0.0 Instruction s^-1
┌ Info: Arithmetic intensity (DRAM)
│   AI_D_DRAM = 10.103490546902368 Instruction Byte^-1
│   AI_F_DRAM = 0.7092428127349074 Instruction Byte^-1
└   AI_H_DRAM = 0.0 Instruction Byte^-1
┌ Info: Arithmetic intensity (L2)
│   AI_D_L2 = 5.165718443517095 Instruction Byte^-1
│   AI_F_L2 = 0.3626220722104721 Instruction Byte^-1
└   AI_H_L2 = 0.0 Instruction Byte^-1
┌ Info: Arithmetic intensity (L1)
│   AI_D_L1 = 2.4958553124543723 Instruction Byte^-1
│   AI_F_L1 = 0.17520355304605265 Instruction Byte^-1
└   AI_H_L1 = 0.0 Instruction Byte^-1

So we get 172GFLOP/s double and 12GFLOP/s single. To be noted this is FLOP/s averaged across all kernels.
The peak FLOP/s is likely higher.