Improve number parsing in streams: use a PushVector to work around slow push! in Base

[KristofferC]

Edit: Rebased on top of #265 which perhaps made this PR less urgent.

This improves the performance of parsing numbers by avoiding the use of push! which is slow (JuliaLang/julia#24909) . Using the benchmark

using JSON
using BenchmarkTools

a = string("[", join([rand(Int) for i in 1:10^3], ","), "]");
@btime JSON.parse($a)

we have before:

julia> @btime JSON.parse(IOBuffer($a))
  318.252 μs (1018 allocations: 32.45 KiB)

and after:

julia> @btime JSON.parse(IOBuffer($a))
  242.448 μs (1015 allocations: 32.38 KiB)

[TotalVerb]

Does this still give improvement after #265?

[KristofferC]

Yes, but only for the StreamingParser case (since after #265, the MemoryParser doesn't use this at all).