optimize int4 gemv kernel with cuda

[yufenglee]

Description

optimize gemv kernel:

1. unroll reduction to improve memory bandwidth
2. leverage 4bits to float16 tricks to save instrutions

m	n	k	symmetric	latency before(us)	latency after(us)
1	4096	4096	TRUE	15.54	8.82
1	4096	4096	FALSE	15.84	9.89
1	4096	11008	TRUE	42.44	19.4
1	4096	11008	FALSE	44.42	21.48
1	11008	4096	TRUE	34.65	17.46
1	11008	4096	FALSE	35.76	20.87
1	12288	4096	TRUE	39.27	19.73
1	12288	4096	FALSE	40.91	25.2
1	22016	4096	TRUE	65.78	38.81
1	22016	4096	FALSE	67.98	48.36