CSV.write internals overhaul

[quinnj]

This PR starts overhauling CSV.write with an aim at slightly better extensibility, simplicity, and most of all, performance. It utilizes the new ryu algorithm for float writing. I haven't quite worked out all the performance kinks yet, but so far, on a 10,000 x 20 Float64 DataFrame, I see the follow differences:

current master:

julia> @time CSV.write(io, df)
  0.064680 seconds (400.10 k allocations: 16.215 MiB, 6.68% gc time)
IOBuffer(data=UInt8[...], readable=true, writable=true, seekable=true, append=false, size=3853698, maxsize=Inf, ptr=3853699, mark=-1)

this PR:

julia> @time CSV.write(io, df)
  0.016382 seconds (10.04 k allocations: 7.983 MiB)
IOBuffer(data=UInt8[...], readable=true, writable=true, seekable=true, append=false, size=3853698, maxsize=Inf, ptr=3853699, mark=-1)

So about a 5x speedup. We're still getting an allocation per row here, so I need to track that down.

[codecov]

Codecov Report

Merging #497 into master will decrease coverage by 1.08%.
The diff coverage is 85.71%.

@@            Coverage Diff             @@
##           master     #497      +/-   ##
==========================================
- Coverage   84.43%   83.34%   -1.09%     
==========================================
  Files           7        7              
  Lines        1195     1243      +48     
==========================================
+ Hits         1009     1036      +27     
- Misses        186      207      +21

Impacted Files	Coverage Δ
src/CSV.jl	78.68% <ø> (-0.06%)	⬇️
src/write.jl	85.14% <85.71%> (-8.97%)	⬇️

---

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 8a6f5bb...def5671. Read the comment docs.