Improvements to attribution

[whitead]

Following discussion from #79, #75, and #78 I've revised the attribution map to track both input and output token indices. As this made the attributions get a bit complicated, I added two dataclasses Attribution and AttributionMap to help make it clear what the meaning is of the various elements of the attribution. One (maybe) disadvantage is that it can be hard to know what the index means because it comes from the tokenizer within the library. One possible change would be to make index be strictly ascending or switch to string index.

Here are examples:

Encoder

s, am = sf.encoder('C1([O-])C=CC=C1Cl.CCO', attribute=True)
print(s)
for a in am:
    print(a)

[C][Branch1][C][O-1][C][=C][C][=C][Ring1][=Branch1][Cl].[C][C][O]
AttributionMap(index=0, token='[C]', attribution=[Attribution(index=0, token='C')])
AttributionMap(index=1, token='[O-1]', attribution=[Attribution(index=3, token='[O-]')])
AttributionMap(index=1, token='[Branch1]', attribution=[Attribution(index=3, token='[O-]')])
AttributionMap(index=2, token='[C]', attribution=[Attribution(index=3, token='[O-]')])
AttributionMap(index=4, token='[C]', attribution=[Attribution(index=5, token='C')])
AttributionMap(index=5, token='[=C]', attribution=[Attribution(index=7, token='C')])
AttributionMap(index=6, token='[C]', attribution=[Attribution(index=8, token='C')])
AttributionMap(index=7, token='[=C]', attribution=[Attribution(index=10, token='C')])
AttributionMap(index=8, token='[Ring1]', attribution=None)
AttributionMap(index=9, token='[=Branch1]', attribution=None)
AttributionMap(index=10, token='[Cl]', attribution=[Attribution(index=12, token='Cl')])
AttributionMap(index=11, token='[C]', attribution=[Attribution(index=13, token='C')])
AttributionMap(index=12, token='[C]', attribution=[Attribution(index=14, token='C')])
AttributionMap(index=13, token='[O]', attribution=[Attribution(index=15, token='O')])

Decoder

s, am = sf.decoder('[C][Branch1][C][O-1][C][=C][C][=C][Ring1][=Branch1][Cl]', attribute=True)
print(s)
for a in am:
    print(a)

C1([O-1])C=CC=C1Cl.CCO
AttributionMap(index=0, token='C', attribution=[Attribution(index=0, token='[C]')])
AttributionMap(index=5, token='[O-1]', attribution=[Attribution(index=1, token='[Branch1]'), Attribution(index=3, token='[O-1]')])
AttributionMap(index=8, token='C', attribution=[Attribution(index=4, token='[C]')])
AttributionMap(index=9, token='=', attribution=[Attribution(index=5, token='[=C]')])
AttributionMap(index=10, token='C', attribution=[Attribution(index=5, token='[=C]')])
AttributionMap(index=12, token='C', attribution=[Attribution(index=6, token='[C]')])
AttributionMap(index=13, token='=', attribution=[Attribution(index=7, token='[=C]')])
AttributionMap(index=14, token='C', attribution=[Attribution(index=7, token='[=C]')])
AttributionMap(index=18, token='Cl', attribution=[Attribution(index=10, token='[Cl]')])
AttributionMap(index=19, token='C', attribution=[Attribution(index=11, token='[C]')])
AttributionMap(index=21, token='C', attribution=[Attribution(index=12, token='[C]')])
AttributionMap(index=23, token='O', attribution=[Attribution(index=13, token='[O]')])

[whitead]

It seems the tox environment is set to test python 3.6 only, despite CI saying its testing python 3.7. This PR uses features from 3.7. Should I rewrite to target 3.6 or was this a misconfiguration in tox env?