Implementation adds non-extension bfloat assignment operators

[steffenlarsen]

As noted in https://github.com/intel/llvm/pull/12477/files#r1471279589, the current implementation of +=, -=, *= and /= add versions taking a T operand on either side of the operator that is convertible to float. Since these are not part of the extension, we have two options:

1. Deprecate them and remove them in a later version.
2. Add them to the extension.

[steffenlarsen]

Tag @rolandschulz, @Pennycook, @AerialMantis, @gmlueck .

[rolandschulz]

I think we should remove all templated operators.

[rolandschulz]

• bfloat16 and half should behave the same
• Half is underspecified (https://gitlab.khronos.org/sycl/Specification/-/issues/455)
• They should align with extended float types (https://wg21.link/p1467) as much as possible
• Current difference is conversion rank (return type of operators with different types), narrowing conversion, and literals
• Fixing those differences is a breaking change.
• For both it might be best if the fixed version goes into a KHR so that the user can opt into the fixed version.