Rationale behind float16 and scipy pinning

[ivustianiu-illumina]

It seems that using float16 to store sparse TF-IDF vectors in entity linkers is the only reason for enforcing old scipy version which prevents package upgrades to modern Python if the linkers are needed.

But what is the rationale behind this except taking less space because it seems to be loaded back as float32? I assume many production use cases are OK with float32 so maybe we can select type based on the scipy version instead of hard limiting the version and raising an exception? I guess there are probably some backward compatibility considerations for keeping float16?

[dakinggg]

The code that requires a specific scipy version should only be relevant if you are creating a linker. Just using it should be fine with any scipy version. Let me know if you're seeing that not work. And yes, the reason is just storage space and download bandwidth.

I'm happy to accept a PR allowing a fallback to float32 in that code section though.

[Odessit007]

Hi @dakinggg. It's my other account, I got a bit overloaded last week so had a chance to have a look only now.

I created a tiny PR, please have a look if it makes sense.
#556