Add splitk hgemm kernel

[xytpai]

Motivation

For improving latency bound hgemm with slim shapes

Technical Details

Use split-k + clean policy

Test Result

Platform: MI308

m	n	k	tile-config	dtype	Torch-hipblas (us)	FlyDSL (us)	Speedup
32	8192	2048	32x256x128-spk4	bf16	32.7	20.2	1.6
32	384	7168	16x128x128-spk8	bf16	12.9	8.4	1.5
32	1024	7168	16x128x128-spk8	bf16	19.1	12.4	1.5
64	1024	7168	32x128x128-spk8	bf16	23.5	15.9	1.5
128	1024	7168	64x128x128-spk8	bf16	35.9	23.4	1.5
256	1024	7168	64x256x64-spk4	bf16	47.5	41.6	1.1

Platform: MI355

m	n	k	tile-config	dtype	Torch-hipblas (us)	FlyDSL (us)	Speedup
32	8192	2048	32x256x128-spk4	bf16	13.2	12.5	1.1
32	384	7168	16x128x128-spk8	bf16	10.9	7.0	1.5
32	1024	7168	16x128x128-spk8	bf16	12.8	10.1	1.3
64	1024	7168	32x128x128-spk8	bf16	13.3	10.7	1.2
128	1024	7168	64x128x128-spk8	bf16	18.1	14.0	1.3
256	1024	7168	64x128x128-spk8	bf16	17.8	16.8	1.1

[coderfeli]

bench mark and test results? also compare large shapes like 8192^3 ?

[xytpai]

bench mark and test results? also compare large shapes like 8192^3 ?

For now, the priority is to cover the kimi cases and adding communication-compute fusion. We can iterate on optimizations in the next step.