Optimizes puff_advantage_kernel using float4 vectorized loads/stores

[jonahsamost]

Config	Elements	Original	New	Speedup
128x64 (small batch)	8,192	0.0383ms	0.0081ms	4.71x
512x64 (medium batch)	32,768	0.0756ms	0.0081ms	9.28x
1024x64 (large batch)	65,536	0.0759ms	0.0082ms	9.28x
2048x64 (xlarge batch)	131,072	0.0766ms	0.0086ms	8.95x
512x16 (short horizon)	8,192	0.0095ms	0.0061ms	1.56x
512x32 (medium horizon)	16,384	0.0379ms	0.0062ms	6.15x
512x128 (long horizon)	65,536	0.1524ms	0.0145ms	10.54x
512x256 (very long horizon)	131,072	0.3030ms	0.0270ms	11.22x
4096x64 (many rows)	262,144	0.0767ms	0.0126ms	6.07x
256x512 (very long horizon)	131,072	0.5980ms	0.0576ms	10.39x

Verify with python and cuda code in tests/ with

nvcc -O3 -arch=sm_86 -shared -Xcompiler -fPIC puff_advantage_standalone.cu -o libpuff_advantage.so
python test_puff_advantage_standalone.py