This PR implements two optimizations to the representation of evidences that significantly decrease memory usage when manipulating large sets of INDRA Statements. The bulk of memory used by INDRA Statements is attributable to the Evidence objects (incl. evidence text) that are attached to them. One approach to decrease memory usage is to define the __slots__ attribute of Evidence to make sure the set of attributes it can have is pre-defined (rather than variable via a __dict__ attribute). This seemed to make a minor difference in memory usage. Much larger memory savings can be achieved if lists of Evidences attached to a Statement are stored in a serialized, compressed form, and only decompressed and deserialized when being accessed. Based on some experiments, a Statement with 100 pieces of Evidence uses 75% less memory using this PR. On some large assembled corpora that I tried, which have Statements with a mixture of number of Evidences, 80% lower memory usage is typical.

Not much of this affects the way INDRA Statements are used, however there is one important difference: when accessing a Statement's evidence (i.e., stmt.evidence) one gets a view of the list evidences rather than a reference to them. So directly manipulating stmt.evidence will not result in persistent changes to the Statement. Rather, one has to do something like:

evs = stmt.evidence
for ev in evs:
    # Make some changes to each ev object
stmt.evidence = evs

to make changes to a Statement's list of Evidences. Some specialized code dealing with Evidence manipulation, as well as some tests needed to be updated. I am still ambivalent about whether this change will cause confusion later, and therefore not sure yet if this PR should be merged.