Support for first and last aggregators - string columns

[Hind-M]

Fix #1105

[alexowens90]

Can we add some tests with None and NaN values in the aggregation columns?
Can we also make the *_with_append tests a bit more complicated, possibly using hypothesis and lmdb_version_store_tiny_segment?

[alexowens90]

This will be prohibitively expensive for large dataframes with fixed-width strings that have not been aggregated, and so false would be sufficient. You need to pass the information in to this method of what this parameter should be.

[alexowens90]

A hacky solution would be to set the orig_type in the aggregator/clause, so that was_coerced_from_dynamic_to_fixed is true in this case. Then add a comment that this is a big hack. Long term we want all reads to go through the ComponentManager, at which point it will decide what to do, so this is OK for now.

[Hind-M]

I first tried the hacky solution, but it's just bringing more complexity since the final dataframe to be compared with the expected one in the tests turns out to have different indexes and data regarding encoding, which involves additional hacking to do.
I think the first solution is fair enough (see this commit).

[Hind-M]

Can we add some tests with None and NaN values in the aggregation columns?

I added None values to the tests (cf. this commit). Regarding NaN values, the tests were already including them, or are you thinking about something specific?

Can we also make the *_with_append tests a bit more complicated, possibly using hypothesis and lmdb_version_store_tiny_segment?

Using hypothesis in *_with_append tests gives an unexpected output when the given input dataframes are the following:
df1:

          grouping_column      a
0               0             0.0

df2:

          grouping_column      a
0               0             0.0

df3:

           grouping_column      a
0               0              0.0
1              00              0.0

using:

lib.write(symbol, df1)
lib.append(symbol, df2)
lib.append(symbol, df3)

Outputs are:

expected_df:                 
grouping_column     a
0                  0.0
00                 0.0

actual_dataframe:  
grouping_column     a
0                  0.0
0                  0.0
00                 0.0

I tried replicating this behavior with a basic test without using hypothesis, but it does give the right expected output dataframe. It seems that something is happening in the PartitionClause on the repartition level which makes it behave this way. I'm not sure what that could be yet...

[Hind-M]

Update: For some reason, all the values in the grouping_column corresponding to 0 don't get to be in the same bucket...