[gluon] Add fuse sigmoid+mul+gemm_splitk

[apinge]

Perf on MI350

mxn_splitk_2s[M=1,bfloat16] : 8 MFLOP / 10.401 us   0.8 TFLOPS   806.7 GB/s 
mxn_splitk_2s[M=1,bfloat16] : 8 MFLOP / 10.120 us   0.8 TFLOPS   829.1 GB/s 
mxn_splitk_2s[M=1,bfloat16] : 8 MFLOP / 10.921 us   0.8 TFLOPS   768.3 GB/s 
mxn_splitk_2s[M=1,bfloat16] : 8 MFLOP / 10.000 us   0.8 TFLOPS   839.1 GB/s 
mxn_splitk_2s[M=1,bfloat16] : 8 MFLOP / 11.880 us   0.7 TFLOPS   706.3 GB/s 
mxn_splitk_2s[M=1,bfloat16] : 8 MFLOP / 9.880 us   0.8 TFLOPS   849.3 GB/s 
mxn_splitk_2s[M=1,bfloat16] : 8 MFLOP / 10.281 us   0.8 TFLOPS   816.1 GB/s 
mxn_splitk_2s[M=1,bfloat16] : 8 MFLOP / 9.880 us   0.8 TFLOPS   849.3 GB/s 
mxn_splitk_2s[M=1,bfloat16] : 8 MFLOP / 9.960 us   0.8 TFLOPS   842.4 GB/s 
mxn_splitk_2s[M=1,bfloat16] : 8 MFLOP / 10.121 us   0.8 TFLOPS   829.0 GB/s 
mxn_splitk_2s[M=1 weight_type=torch.bfloat16] acc OK
final selected TILE_M=16, TILE_N=32
mxn_splitk_2s[M=4,bfloat16] : 34 MFLOP / 8.840 us   3.8 TFLOPS   949.9 GB/s 
mxn_splitk_2s[M=4,bfloat16] : 34 MFLOP / 10.160 us   3.3 TFLOPS   826.5 GB/s 
mxn_splitk_2s[M=4,bfloat16] : 34 MFLOP / 10.200 us   3.3 TFLOPS   823.2 GB/s 
mxn_splitk_2s[M=4,bfloat16] : 34 MFLOP / 10.321 us   3.3 TFLOPS   813.6 GB/s 
mxn_splitk_2s[M=4,bfloat16] : 34 MFLOP / 10.000 us   3.4 TFLOPS   839.7 GB/s 
mxn_splitk_2s[M=4,bfloat16] : 34 MFLOP / 13.440 us   2.5 TFLOPS   624.8 GB/s 
mxn_splitk_2s[M=4,bfloat16] : 34 MFLOP / 11.000 us   3.1 TFLOPS   763.3 GB/s 
mxn_splitk_2s[M=4,bfloat16] : 34 MFLOP / 11.041 us   3.0 TFLOPS   760.5 GB/s 
mxn_splitk_2s[M=4,bfloat16] : 34 MFLOP / 10.960 us   3.1 TFLOPS   766.1 GB/s 
mxn_splitk_2s[M=4,bfloat16] : 34 MFLOP / 11.240 us   3.0 TFLOPS   747.0 GB/s 
mxn_splitk_2s[M=4 weight_type=torch.bfloat16] acc OK
final selected TILE_M=16, TILE_N=32
mxn_splitk_2s[M=16,bfloat16] : 134 MFLOP / 8.960 us   15.0 TFLOPS   939.9 GB/s 
mxn_splitk_2s[M=16,bfloat16] : 134 MFLOP / 10.200 us   13.2 TFLOPS   825.6 GB/s 
mxn_splitk_2s[M=16,bfloat16] : 134 MFLOP / 10.441 us   12.9 TFLOPS   806.6 GB/s 
mxn_splitk_2s[M=16,bfloat16] : 134 MFLOP / 10.000 us   13.4 TFLOPS   842.1 GB/s 
mxn_splitk_2s[M=16,bfloat16] : 134 MFLOP / 10.560 us   12.7 TFLOPS   797.5 GB/s 
mxn_splitk_2s[M=16,bfloat16] : 134 MFLOP / 10.640 us   12.6 TFLOPS   791.5 GB/s 
mxn_splitk_2s[M=16,bfloat16] : 134 MFLOP / 10.961 us   12.2 TFLOPS   768.3 GB/s 
mxn_splitk_2s[M=16,bfloat16] : 134 MFLOP / 10.600 us   12.7 TFLOPS   794.5 GB/s 
mxn_splitk_2s[M=16,bfloat16] : 134 MFLOP / 10.920 us   12.3 TFLOPS   771.2 GB/s 
mxn_splitk_2s[M=16,bfloat16] : 134 MFLOP / 10.880 us   12.3 TFLOPS   774.0 GB/s 
mxn_splitk_2s[M=16 weight_type=torch.bfloat16] acc OK
dict_tile_mn {'1': (16, 32), '4': (16, 32), '16': (16, 32)}

summary:
mxn_splitk_2s[torch.bfloat16 B=1   (16x32)]:    10 us,  814.3 GB/s,  0.8 tflops
mxn_splitk_2s[torch.bfloat16 B=4   (16x32)]:    11 us,  773.9 GB/s,  3.1 tflops
mxn_splitk_2s[torch.bfloat16 B=16  (16x32)]:    11 us,  796.8 GB/s, 12.7 tflops