Fix handling of hash collisions in row_group_slots

[nalimilan]

Original code was correct but JuliaData/DataTables.jl#79 incorrectly moved break outside of the if.

Fixes #1978.

[bkamins]

Thank you for a quick fix of this. I have checked the logic, and it seems OK so I approve 😄.

@mattBrzezinski - it would be great if you had a chance to test it.

@nalimilan If I understand it correctly df = DataFrame(A=1:2_000_000) test by @mattBrzezinski already exposes the collisions and I think this should be small enough to be added to our tests.

[mattBrzezinski]

Good morning.

I can verify this fixes the issue later today, and write some unit tests for it as well!

@nalimilan would you like me to push commits to this branch? Or just create a new one for the unit-tests?

[nalimilan]

I've added a test: can you just check that it reproduces the problem on 32-bit without the patch? Thanks!

[bkamins]

Now I have four Julia versions installed in parallel to be able to develop DataFrames.jl 😄.

The test correctly reproduces the error prior to the change in this PR:

julia> gd = groupby(df, :A)
ERROR: ReadOnlyMemoryError()
Stacktrace:
 [1] getindex at .\array.jl:728 [inlined]
 [2] group_rows(::DataFrame, ::Bool, ::Bool, ::Bool)
 [3] #groupby#157(::Bool, ::Bool, ::typeof(groupby), ::DataFrame, ::Array{Symbol,1})
 [4] #groupby at .\none:0 [inlined]
 [5] #groupby#158 [inlined]
 [6] groupby(::DataFrame, ::Symbol)
 [7] top-level scope at REPL[6]:1

[nalimilan]

I've added some more checks to make sure that even if no error happens for some reason, we don't mix two different values that hash equal in the same group.