Improve canonicalization performance

[ddeschepper]

We're noticing big performance issues when using longturtle serialization on some graphs. I've been able to narrow this down to the performance of canonicalization, which is also tracked in issue #2528.

Looking into it I found that the current implementation of the _traces method of _TripleCanonicalizer causes much of the performance impact.

This PR reduces the complexity of _traces, which leads to a performance gain of at least an order of magnitude in our worst cases (100s -> 4s). All rdflib tests still pass, and additionally, I've tested these changes with our set of a few hundred examples that are longturtle serialized, which causes no changes in the serialization output.

The author of the linked issue has created a performance test that, with the current code, gives the following results on my machine:

file: test1.ttl
isomorphic: 0.07787537574768066
canonical: 0.03909921646118164

file: test2.ttl
isomorphic: 3.528538942337036
canonical: 1.8337273597717285

file: test3.ttl
isomorphic: 20.140648365020752
canonical: 9.535402774810791

where my new version results in:

file: test1.ttl
isomorphic: 0.012566566467285156
canonical: 0.006159543991088867

file: test2.ttl
isomorphic: 0.15960264205932617
canonical: 0.09874987602233887

file: test3.ttl
isomorphic: 0.531768798828125
canonical: 0.2606654167175293

[edmondchuc]

The right-hand side of the or is never evaluated as best_score is None is always true here. Can you please review this.