Issues when increasing L1 K tile size

[yzhang93]

I have observed several issues when increasing the L1 K tile size from 32 to 64.

1. Vectorization path slows down significantly (3ms -> 10ms) with default settings, i.e., peano flag O3 and function outlining.
2. Vectorization path has program memory issue if function outlining is disabled.
3. Ukernel path has error ld.lld: error: undefined symbol: matmul_bf16_bf16_f32_32x32x64_4x8x4 with function outlining enabled.
4. Ukernel path doesn't have the above error if function outlining is disabled, however the performance doesn't change.

CC: @newling @jtuyls