fix fp8 blockscale configs for deepseek

[juuso-oskari]

This PR provides finetuned configs for deepseek fp8 blockscale. Also fixes the benching script (K = intermediate dimension // 2 for fc2 if GLU is used, NOT N = intemediate dimension * 2 for fc1).

[anhminhnguyenhoang]

Tensor shapes M=[64, 128, 256], N=2112, M=7168 for Deepseek model has been tuned for Triton block configs facilitating at least 2x uplift in performance in comparison to the current state in main branch.

# before
     M     N     K  throughput (TFLOPs)  time (ms)  bandwidth (GB/s)
0   64  2112  7168            14.984021   0.128799        123.609343
1  128  2112  7168            30.529371   0.127317        131.801976
2  256  2112  7168            66.172665   0.117496        154.754403

# after
     M     N     K  throughput (TFLOPs)  time (ms)  bandwidth (GB/s)
0   64  2112  7168            27.073608   0.068483        221.219118
1  128  2112  7168            89.067354   0.043015        372.392590
2  256  2112  7168           180.079694   0.045219        414.613027

Ps: Results rendered on asrock-1w300-e0-3.mkm.dcgpu (mi350xas2)