Add Bfloat16: alternative 16 bit floating point precision to half

[yhmtsai]

This PR adds bfloat16 precision but user can only select one of bfloat16 or half in ginkgo now.
There will be another pr to enable the possibility of having both at the same time, which should mostly contain the convert function/precision chain and some improvement to reduce the copy-paste stuff.

This PR currently use gko::float16 for either half or bfloat16. I think float16 is confusing to people because it is the same term used by half.

summary for different vendor bfloat16,

• CUDA:
  • they use __nv_bfloat16
  • some ptx requires sm_80 to support (not have the guard in this pr). normal operations require at least sm_80 or cuda 12.2 similar to half
  • only use b16 in .reg, which is unlike f16 for half precision
• HIP:
  • they have two bfloat16 format (hip_bfloat16 and __hip_bfloat16), hip_bfloat16 is quite early support but likely use float for arithmetic operation internally. __hip_bfloat16 has more native operation on bfloat16. __hip_bfloat16 supports from 5.6.0 but we need at least 6.2.0 to get enough implementation for the operation overload and conversion. before 5.4.0, it does not contain operator=(float).
  • more trouble on finding sqrt. hip tries to use system sqrt only in lambda function in the kernel (works in __global__). I need to add __device__ I guess it limits the searching space because we only provide the sqrt(bfloat16) in device.
• SYCL:
  • it is in the experimental namespace
  • not proper implementation or they have some default implementation for std::numeric_limits on bfloat16 because I do not get any compilation issue. provide device_numeric_limits now
  • unary operation - on rvalue will gives float before 2025.0.1 because it only accepted non-const reference before 2025.0.1

[MarcelKoch]

I have mostly smaller comments for this, rest looks good.

[sonarqubecloud]

Quality Gate passed

Issues
32 New issues
0 Accepted issues

Measures
0 Security Hotspots
73.2% Coverage on New Code
14.7% Duplication on New Code

See analysis details on SonarQube Cloud