BUG: Fix rebuild regression

[larsoner]

On fc8247a (good) or earlier or this PR, doing two consecutive builds in a row I see:

$ make clean && make html_dev-noplot && make html_dev-noplot
building [html]: targets for 50 source files that are out of date
updating environment: [new config] 755 added, 0 changed, 0 removed
...
building [html]: targets for 0 source files that are out of date
updating environment: 0 added, 7 changed, 0 removed

On #8905 / 102393f (bad) or master, I see:

$ make clean && make html_dev-noplot && make html_dev-noplot
building [html]: targets for 50 source files that are out of date
updating environment: [new config] 755 added, 0 changed, 0 removed
...
building [html]: targets for 755 source files that are out of date
updating environment: 0 added, 7 changed, 0 removed

This regression causes all documents to be rewritten in the writing output... stage, which takes the incremental build from ~10 seconds to ~2 minutes, which is a tough efficiency regression when trying to iteratively write docs.

I think this is due to object() not having a stable repr across runs because the memory location / id can change, and id is in the __repr__ of object instances. So this PR ensures that UNSET has a stable __repr__, and fixes my incremental doc builds.

I went to add a test and realized it's not so easy because you actually need multiple Python processes to get this effect, because within a given process the UNSET will have a stable repr. So that's why there is no test. And maybe this is why it escaped existing tests if there are any that check to make sure things are not unnecessarily rebuilt.

[tk0miya]

Thanks!