Add optimized implementation of fill for types supported by memset.

[maleadt]

Before: our broadcasting fill! is slower than a memory copy...

julia> @btime cuzeros(Float32,10,10)
  6.552 μs (47 allocations: 1.61 KiB)

julia> @btime CuArray(zeros(Float32,10,10))
  4.616 μs (18 allocations: 1.94 KiB)

After:

julia> @btime cuzeros(Float32,10,10)
  2.162 μs (15 allocations: 352 bytes)

Would be nice if we could dispatch on haskey(MemsetTypes, sizeof(T)) instead of enumerating like now.

[codecov-io]

Codecov Report

Merging #178 into master will increase coverage by 0.07%.
The diff coverage is 100%.

@@            Coverage Diff             @@
##           master     #178      +/-   ##
==========================================
+ Coverage   43.47%   43.55%   +0.07%     
==========================================
  Files          34       34              
  Lines        2123     2126       +3     
==========================================
+ Hits          923      926       +3     
  Misses       1200     1200

Impacted Files	Coverage Δ
src/array.jl	51.35% <100%> (+2.05%)	⬆️

---

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 599942e...832db73. Read the comment docs.

[maleadt]

Probably should implement this with a constructor instead of cuzeros, ref. JuliaLang/julia#11557 ?

[maleadt]

Let's merge this since it's a strict improvement.