Speedup Lut-gemm for issue #30

[5000user5000]

Based on the investigation in ISSUE #30 , it was found that LUT GEMM suffers from high L1 cache miss rates, making memory latency the performance bottleneck rather than computation. Several optimizations were applied, such as prefetching the LUT into L1 cache and tiling the matrices to avoid loading overly large chunks that cause cache misses. Additionally, multithreading was introduced to further improve performance. The original naive implementation was also updated to support multithreading, making it easier to compare performance under a fixed thread count.
Lastly, tests/test_matrix_ops.cpp was added to help dump the assembly of both the naive (float) and LUT GEMM versions for comparison.

[5000user5000]

@yungyuc
I've already looked into Monday's request — identifying why the LUT implementation performs similarly to, or even slightly worse than, the naive version. In this PR, I've also made efforts to mitigate some of LUT's drawbacks, such as excessive L1 cache misses.

[yungyuc]

Good findings. When adding your findings into the presentation, make sure you address all the assessing points for grading.

Where in the code you want my review? Please make inline annotation.

Don't hesitate to merge the PR. Review and discussions still work after merging.