Optimize pow5_12(::Float32)

[kimikage]

Without PR #482, pow5_12(::Float32) was never called because linear RGB values are promoted to Float64 in the transform matrix multiplication.
The new pow5_12(::Float32) has an accuracy of about 0.5 ULP.

[codecov]

Codecov Report

Merging #493 (c4ae4f9) into master (9710272) will increase coverage by 0.06%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master     #493      +/-   ##
==========================================
+ Coverage   93.40%   93.47%   +0.06%     
==========================================
  Files           9        9              
  Lines        1061     1072      +11     
==========================================
+ Hits          991     1002      +11     
  Misses         70       70              

Impacted Files	Coverage Δ
src/conversions.jl	99.38% <100.00%> (+<0.01%)	⬆️
src/utilities.jl	98.93% <100.00%> (+0.06%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 9710272...c4ae4f9. Read the comment docs.

[kimikage]

Benchmark

cf. #482 (comment)

julia> @btime convert.(RGB, $xyz_f32);
  36.715 ms (2 allocations: 11.44 MiB) # before (PR #482)
  31.418 ms (2 allocations: 11.44 MiB) # after

julia> @btime convert.(RGB{N0f8}, $xyz_f32);
  42.136 ms (2 allocations: 2.86 MiB) # before (PR #482)
  40.123 ms (2 allocations: 2.86 MiB) # after

[kimikage]

The connection between the curve and the line in fxyz2lab exactly matches in value and slope. Therefore, it is possible to merge the two with min. On the other hand, in srgb_compand there is a small gap between the curve and the line.

julia> Float64(big"0.0031308"^(5/big"12") * big"1.055" - big"0.055")
0.04044990748269015

julia> Float64(big"0.0031308" * big"12.92") # != 0.04045
0.040449936