Update default M from 2 to 8

[a3213105]

So when matrix M<=8, we will use ig kernel to boost bf16 gemms, otherwise use oneDNN kernel.
From the performance data， we can see that with M=8， It can boost next token performance of BS=4/8, especially for BS=4.
This patch doesn't affect first token latency since the M of first latency is larger than 8.

batchsize	1	2	4	8	16
chatglm-6b
original M=2	30.58	31.03	63.56	67.61	78.38
M=7	29.54	32.58	41.28	68.14	78.2
M=8	29.87	32.88	40.57	66.82	78.83
chatglm2-6b
original M=2	30.7	31.61	63.11	67.32	83.61
M=7	29.94	33.3	41.1	70.68	83.49
M=8	31.76	31.84	41.01	63.84	78.83
llama-7b
original M=2	32.38	36.47	71.6	78.45	92.83
M=7	33.88	37.54	46.43	79.64	93.27
M=8	32.64	36.59	46.04	75.14	93.18