Split K for SharedMoE fusion

[jianan-gu]

Performance on a GNR_DDR ( so far )
Current splitk num is 4

Prefill (total/avg/layers) 1K	BS1	BS4
Baseline	105.897ms/1.826ms/58	396.073ms/6.829ms/58
Splitk=2	100.539ms/1.733ms/58	373.898ms/6.447ms/58
Splitk=4	102.118ms/1.761ms/ 58	354.326ms/6.109ms/58
Splitk=8	107.627ms/1.856ms/58	358.720ms/6.185ms/58

Decode (total/avg/layers) 1K	BS1	BS4
Baseline	4.904ms/84.549us/58	5.270ms/90.864us/58
Splitk=2	4.491ms/77.430us/58	5.524ms/95.247us/58
Splitk=4	4.381ms/75.539us/58	5.001ms/86.217us/58
Splitk=8	4.532ms/78.137us/58	5.079ms/87.572us/58

DS R1 test on GNR_MCR

Reference
BS4
First 
sgl-kernel::shared_expert_cpu         4.68%     337.369ms         4.69%     337.822ms       5.825ms            58 
Next 
sgl-kernel::shared_expert_cpu         6.31%       4.541ms         6.44%       4.633ms      79.880us            58  
BS1
First 
sgl-kernel::shared_expert_cpu         5.24%      92.125ms         5.25%      92.237ms       1.590ms            58
Next 
sgl-kernel::shared_expert_cpu         6.61%       3.720ms         6.68%       3.763ms      64.884us            58


Improved
BS4
First 
sgl-kernel::shared_expert_cpu         4.46%     272.729ms         4.46%     273.136ms       4.709ms            58  
Next  
sgl-kernel::shared_expert_cpu         5.72%       4.086ms         5.87%       4.189ms      72.229us            58              

BS1
First 
 sgl-kernel::shared_expert_cpu         5.08%      87.134ms         5.09%      87.346ms       1.506ms            58
Next 
sgl-kernel::shared_expert_cpu         6.33%       3.544ms         6.49%       3.636ms      62.695us            58
    

Acc test cases all pass

Python https://github.com/mingfeima/sgl-cpu-tests/blob/main/test_shared_experts.py

run_single_test(1, 704, 7168, 1, torch.bfloat16)
run_single_test(1, 1024, 1024, 1, torch.bfloat16)
run_single_test(4, 704, 7168, 1, torch.bfloat16)
run_single_test(4, 1024, 1024, 1, torch.bfloat16)
run_single_test(128, 704, 7168, 1, torch.bfloat16)
run_single_test(128, 1024, 1024, 1, torch.bfloat16)