Add shift by multiple constant

[AntoinePrv]

@serge-sans-paille @JohanMabille this ideas works, but I cannot figure out how to refactor bitwise_lshift_as_twice_larger into a separate header.

The issue:

• xsimd_sse2.hpp needs utils/shits.hpp for bitwise_lshift_as_twice_larger
• but bitwise_lshift_as_twice_larger needs definitions from xsimd_sse2.hpp.
• and likewise for avx2

Forward declaring all the needed functions overloads from sse2 and avx2 and the arch types in utils/shits.hpp.
Perhaps it should use the functions from xsimd_api instead? (which would also be better if there is a better implementation further down in the inheritance tree).

Close #1218

[AntoinePrv]

Also we need to add all the proper forward to dynamic for all other architectures to avoid ending up with a common implementation (unless that's all there is).

[AntoinePrv]

I named this util as common seems to be more used for implementing the common architecture.

[AntoinePrv]

Probably should remove R req here

[serge-sans-paille]

@AntoinePrv unless there's is a huge need on your side, I won't schedule that one for the release. Let's just add masked load / store support for other architecture so that we can release, then begin a new commit wave :-)

[AntoinePrv]

@serge-sans-paille I had hoped I could do it quickly before the weekend but it seems it will take longer.
Let's release indeed, I'll add the fix in Arrow with a TODO for cleaning it up in a future release.

[AntoinePrv]

I have a few questions on the current implementation

• For the public API, wouldn't this benefit from a std::enable_if<std::is_integral<T>::value>::type?
  

  xsimd/include/xsimd/types/xsimd_api.hpp

  Line 370 in 23faec4

  template <class T, class A>

• For neon, the unsigned API takes shifts as signed, yet there is not overload for unsigned shifts, which means that in this case we fallback on common?
  

  xsimd/include/xsimd/arch/xsimd_neon.hpp

  Line 2421 in 23faec4

  XSIMD_INLINE batch<T, A> bitwise_lshift(batch<T, A> const& lhs, batch<as_signed_integer_t<T>, A> const& rhs, requires_arch<neon>) noexcept