Use `IamDataFrame` as return-type of `filter_*()`

After some more dabbling with the current codebase, I suggest using an IamDataFrame as the return-type directly of the filter_*() functions, something along following structure:

def filter_*(*args, **kwargs):
    data = ...  # processing from osemosys to dataframe
    return IamDataFrame(data, **kwargs)

In my opinion, this would have several benefits:

Better performance, because pyam internally uses an indexed pd.Series instead of a pd.DataFrame
Earlier catching of errors (e.g., duplicates as in #17)
pyam was already refactored to use pd.concat internally instead of pd.DataFrame.append (even pyam.IamDataFrame.append uses pd.concat()), so this would solve #14
You could use specific aggregation-methods by pyam directly (e..g, see aggregate()), so
```
aggregated = aggregate(data)

if not aggregated.empty:
    iamc = make_iamc(aggregated, config['model'], config['scenario'], input['iamc_variable'], unit)
    blob.append(iamc)
```
could be simplified to
```
data.aggregate(variable=data.variable, append=True)
```
and you get validation (guarding against duplicates, ...) and better performance (because indexed series) free of charge.
- You could simplify the test suite by using the pyam.testing module, which eliminates the need to set an index, sort, fix data types, etc (see #16).

Happy to discuss here or in person at the ECEMF meeting!

OSeMOSYS / osemosys2iamc

Use `IamDataFrame` as return-type of `filter_*()` #21