Make Each(Drop|Replace|Fail)Null iterators faster

[nalimilan]

The compiler previously did not know that the result would never be Null.
This dramatically improves their efficiency.

Also require that the type of replacement matches eltype of iterator
for EachReplaceNull, which is much faster. Nulls.replace() already
takes care of converting it on construction.

Optimizations need to rely on the particular fields of DataArray.
This requires restricting them to DataArray, and moving code
to dataarray.jl since the type had not been defined in abstractdataarray.jl.

Rather than adding specialized iterators for PooledDataArray, which
is going to be deprecated, rely on the generic iterators provided
by Nulls, which are already faster than the current ones.

Cf. JuliaData/Missings.jl#50

---

Before (on Julia 0.6.0):

Main> using Nulls

Main> using DataArrays

Main> using BenchmarkTools

Main> x = DataArray(Vector{Union{Int, Null}}(rand(Int, 100_000)));

Main> x[rand(1:length(x), 10_000)] = null;

Main> skip_iter(x) = sum(Nulls.skip(x))
skip_iter (generic function with 1 method)

Main> @btime skip_iter(x);
  6.988 ms (552024 allocations: 8.42 MiB)

Main> replace_iter(x, y) = sum(Nulls.replace(x, y))
replace_iter (generic function with 1 method)

Main> @btime replace_iter(x, 0);
  31.703 ms (1051004 allocations: 17.56 MiB)

Main> y = DataArray(Vector{Union{Int, Null}}(rand(Int, 100_000)));

Main> fail_iter(x) = sum(Nulls.fail(x))
fail_iter (generic function with 1 method)

Main> @btime fail_iter(y);
  31.586 ms (1098980 allocations: 18.29 MiB)

After (on Julia 0.6.0):

Main> using Nulls

Main> using DataArrays

Main> using BenchmarkTools

Main> x = DataArray(Vector{Union{Int, Null}}(rand(Int, 100_000)));

Main> x[rand(1:length(x), 10_000)] = null;

Main> skip_iter(x) = sum(Nulls.skip(x))
skip_iter (generic function with 1 method)

Main> @btime skip_iter(x);
  516.099 μs (4 allocations: 64 bytes)

Main> replace_iter(x, y) = sum(Nulls.replace(x, y))
replace_iter (generic function with 1 method)

Main> @btime replace_iter(x, 0);
  676.586 μs (4 allocations: 80 bytes)

Main> y = DataArray(Vector{Union{Int, Null}}(rand(Int, 100_000)));

Main> fail_iter(x) = sum(Nulls.fail(x))
fail_iter (generic function with 1 method)

Main> @btime fail_iter(y);
  526.172 μs (4 allocations: 64 bytes)

[ararslan]

[nalimilan]

OK to merge?