[BUG] [CPU] Some ops fallback to fp32 implementation even fro fp64

[awni]

Some ops (exp, sin, cos, erf, erfinv) use the fp32 approximation for double and are hence not so accurate for double precision. Would be good to have specializations of these for double as well.

[awni]

The fp32 implementations are here: https://github.com/ml-explore/mlx/blob/main/mlx/backend/cpu/simd/math.h

We should probably just add a constexpr for double + double approximations for those.

[Anri-Lombard]

Hi @awni, I can try and take this on if that's helpful?

The plan would be to add if constexpr double branches in math.h for the 4 implementations (exp, sincos, erf, erfinv) using Cephes double-precision coefficients. Also add the missing fp16 wrappers in accelerate_fp16_simd.h.

[awni]

Great, I will assign to you!