sgemm: simplify kernel_x86_avx logic and reduce shuffle overhead

[SongXiaoXi]

This PR simplifies the implementation of the AVX-based 8x8 SGEMM microkernel in kernel_x86_avx, again.

This refactor is motivated by the fact that instructions like moveldup, movehdup, permute, and permute2f128 are all executed exclusively on port 5. On FMA-capable platforms, this creates a bottleneck that the fmadd throughput cannot hide, especially when combined with loads followed by permutes. The new implementation reduces port 5 pressure by avoiding these sequences entirely during the inner loop.

By the way, the theoretical maximum FLOPS kernel on FMA-capable platforms should resemble a 6x16 layout using _mm256_broadcast_ss(from memory) followed by two _mm256_fmadd_ps. However, this comes at the cost of losing the transposition flexibility of square kernels that adapt to the row/column-major layout of C.

Performance

On Intel i7-10700k
Before:

m	k	n	layout	type	average_ns	minimum_ns	median_ns	samples	GFLOPS	nc	kc	mc	threads
384	384	384	FCC	f32	974,601	973,492	974,549	1560	116.1975085188708				0
450	450	450	FCC	f32	1,594,266	1,584,060	1,590,237	1560	114.31592971310936				0
512	512	512	FCC	f32	2,334,892	2,289,947	2,357,830	1560	114.96696892190303				0

After:

m	k	n	layout	type	average_ns	minimum_ns	median_ns	samples	gflops	nc	kc	mc	threads
384	384	384	FCC	f32	892269	891653	892212	1560	126.91935727902684				0
450	450	450	FCC	f32	1460456	1456068	1460988	1560	124.78979168150222				0
512	512	512	FCC	f32	2108191	2101841	2111378	1560	127.32976091824698				0