(Replaces #236)
Base.Math.log1p calculation relies on two procedures , log_proc1 and log_proc2 from the literature Tang (1990). The current implementation upscales to Float64 internally and as such is incompatible with Metal.

See also relevant discussions #234, #236

This PR ports log1pf from openlibm

Error Analysis (ULP) (Ran on CPU)

This implementation v. the original C implementation (Float32)

max 1.0 at x = 3.0777041e38
mean 0.00010909553601139363

This implementation v. Base.Math.log1p(::Float64)

max 0.8388471007347107 at x = 0.40553576
mean 0.11881815022236436

This implementation v. Base.Math.log1p(::Float32)

max 1.0 at x = 3.4005017e38
mean 0.003624081612764254

Error Analysis (ULP) (Ran on GPU using 100,000 points)

v. Base.Math.log1p(::Float64)

max 0.5023159980773926 at x = 2.3837697e38
mean 0.24963651393659414

v. Base.Math.log1p(::Float32)

max 1.0 at x = 3.2846083e38
mean 0.00113