WIP: Fix type instability in groupby()

[nalimilan]

This ensures the loop over rows is type stable, which increases performance
a lot.

This is WIP because it must be balanced with #3. It's mainly intended to allow for performance comparisons.

The same strategy could be applied to DataFrames, which are faster than DataTables currently, but much slower than with this PR.

using DataTables
using BenchmarkTools

dt1 = DataFrame(v1 = PooledDataArray(repeat(1:10, inner=1000)), v2 = PooledDataArray(repeat(1:10, outer=1000)));
dt2 = DataFrame(v1 = PooledDataArray(repeat(1:100, inner=100)), v2 = PooledDataArray(repeat(1:100, outer=100)));
dt3 = hcat(dt1, dt2)

@benchmark groupby(dt1, [:v1, :v2])
@benchmark groupby(dt2, [:v1, :v2])
@benchmark groupby(dt3, [:v1, :v2, :v1_1, :v2_1])


# BEFORE
julia> @benchmark groupby(dt1, [:v1, :v2])

BenchmarkTools.Trial: 
  memory estimate:  1.22 MiB
  allocs estimate:  66536
  --------------
  minimum time:     4.527 ms (0.00% GC)
  median time:      4.747 ms (0.00% GC)
  mean time:        5.047 ms (2.08% GC)
  maximum time:     12.827 ms (18.36% GC)
  --------------
  samples:          988
  evals/sample:     1
  time tolerance:   5.00%
  memory tolerance: 1.00%

julia> @benchmark groupby(dt2, [:v1, :v2])
BenchmarkTools.Trial: 
  memory estimate:  1.99 MiB
  allocs estimate:  85972
  --------------
  minimum time:     5.063 ms (0.00% GC)
  median time:      5.518 ms (0.00% GC)
  mean time:        6.066 ms (3.17% GC)
  maximum time:     13.451 ms (23.09% GC)
  --------------
  samples:          821
  evals/sample:     1
  time tolerance:   5.00%
  memory tolerance: 1.00%

julia> @benchmark groupby(dt3, [:v1, :v2, :v1_1, :v2_1])
BenchmarkTools.Trial: 
  memory estimate:  20.45 MiB
  allocs estimate:  238751
  --------------
  minimum time:     17.675 ms (0.00% GC)
  median time:      20.564 ms (7.95% GC)
  mean time:        20.652 ms (7.97% GC)
  maximum time:     25.163 ms (6.69% GC)
  --------------
  samples:          242
  evals/sample:     1
  time tolerance:   5.00%
  memory tolerance: 1.00%

# AFTER:
julia> @benchmark groupby(dt1, [:v1, :v2])
BenchmarkTools.Trial: 
  memory estimate:  122.64 KiB
  allocs estimate:  21
  --------------
  minimum time:     63.710 μs (0.00% GC)
  median time:      69.333 μs (0.00% GC)
  mean time:        76.896 μs (5.47% GC)
  maximum time:     971.994 μs (91.85% GC)
  --------------
  samples:          10000
  evals/sample:     1
  time tolerance:   5.00%
  memory tolerance: 1.00%

julia> @benchmark groupby(dt2, [:v1, :v2])
BenchmarkTools.Trial: 
  memory estimate:  520.89 KiB
  allocs estimate:  29
  --------------
  minimum time:     148.724 μs (0.00% GC)
  median time:      163.476 μs (0.00% GC)
  mean time:        188.497 μs (8.29% GC)
  maximum time:     1.161 ms (74.62% GC)
  --------------
  samples:          10000
  evals/sample:     1
  time tolerance:   5.00%
  memory tolerance: 1.00%

julia> @benchmark groupby(dt3, [:v1, :v2, :v1_1, :v2_1])
BenchmarkTools.Trial: 
  memory estimate:  16.56 MiB
  allocs estimate:  36
  --------------
  minimum time:     7.442 ms (0.00% GC)
  median time:      8.863 ms (5.16% GC)
  mean time:        8.971 ms (7.86% GC)
  maximum time:     13.647 ms (22.29% GC)
  --------------
  samples:          556
  evals/sample:     1
  time tolerance:   5.00%
  memory tolerance: 1.00%

[nalimilan]

Closing in favor of #17.