Replace BLAS with packed_simd

[generall]

https://github.com/rust-lang/packed_simd

Is your feature request related to a problem? Please describe.
BLAS is currently used for a single purpose - to make dot production faster. It looks like overkill.
Compile time is pretty high and it complicates the development

Describe the solution you'd like
Replace BLAS with packed_simd library, which appear to be much lighter.

Describe alternatives you've considered
Use some other lightweight library for SIMD (idk which one)

Additional context
None

[salamer]

hora have used packed_simd as its SIMD backend, but it can only be built in nightly version, and packed_simd has not iterated for a couple of months, and packed_simd looks like merge into stable version is still in the foreseeable future. if you don't have installation pressure, BLAS may be a better choice, because it has proved itself in many projects.

[generall]

Hi @salamer! I was looking at hora when submitted this issue;) As a matter of fact, BLAS is a pain point when it comes to compilation. I also would like to avoid one unnecessary vector copy which is only used due to ndarray package. I would probably have to spend some more time on this topic and choose fast and at the same time easy to use package for SIMD.

[generall]

Different approach to consider https://github.com/zhao-lang/redis_hnsw/blob/master/src/hnsw/metrics.rs

[generall]

BLAS removed in https://github.com/qdrant/qdrant/releases/tag/v0.7.0