Closes #172

Co-authored-by: Takuya Hashimoto <[email protected]>

…into main-dev

On the Leipzig1M dataset, LibC vs SZ:

~ 128b lines, aligned: 2.3 vs 2.6 GB/s
~ 128b lines, unaligned: 2.34 vs 2.53 GB/s
~ 5b tokens, aligned: 0.1 vs 0.1 GB/s
~ 5b tokens, unaligned: 0.1 vs 0.1 GB/s
~ 124 MB, aligned: 19.6 vs 20.3 GB/s
~ 124 MB, unaligned: 19.6 vs 20.3 GB/s

…into main-dev

Previously SZ would build too many
targets for each debugging session.

This commit accelerates the `sz_fill_avx2`
and `sz_copy_avx2` by avoiding unaligned writes.

It also adds an `sz_equal_avx2` to help validate
large files with matching checksums faster.

It also adds a placeholder for `sz_order_avx2`,
discouraging further optimizations.

C++ API with a matching argument order was added
to mimic `std::memcpy`, `std::memset`, `std::memmove`.

Matching `test_memory_utilities` tests were extended.

In AVX-512, similar to GLibC we should use the
register space to load more data simultaneously
and avoid loops and data-dependency between
iterations.

…into main-dev

The new `sz_look_up_transform` API implements
a 256-byte lookup table using serial code and AVX-512
that can significantly accelerates text and image
processing.

The AVX-512 implementation reaches 18 GB/s on
Intel Sapphire Rapids CPU, while serial code stays
around 3 GB/s for large files.