Proposal for `batch_constant` biwise shift.

[AntoinePrv]

There is a trick borrowed from Lemire that I am using to implement left shift when unavailable: multiply by 2^n. (The rest of the trick is that for right shifting, if you known that you data will "fit", you can left shift everything by some amount followed by a constant right shift on the whole batch).

With this, we can add missing variable shift on x86_64 for:

• SSE2:
  • left uint16_t
  • left uint32_t
• AVX2:
  • left uint16_t

On top of this, we can implement variable shift by calling twice a shift twice + mask on an integer size twice as big, so that further enables:

• SSE2:
  • left uint8_t
• AVX2:
  • left uint8_t
  • right uint16_t
  • right uint8_t

The main drawback is adding a new batch_constant API for left shift (more code to re-dispatch in the general case). This is necessary because we need the left shift to create the power of two.

I think this is still useful because it enables some shifts for the whole SSE inheritance.

I have a working partial implementation that I could turn into a PR in the next couple days, and hopefully into the 14.0 release. What do you think @serge-sans-paille?

[serge-sans-paille]

For left shifting, that should probably be the generic default implementation, right? I'm fine with that either way.