Compute subset errors everywhere

[lqd]

This PR implements what we discussed last sprint:

• make the Opt variant compute subset errors, tracking the subsets of placeholder origins, because placeholder loans don't work in this situation
• make the Naive variant compute subset errors using the "easy subset-filtering" because we compute the full subset transitive closure here
• do nothing for now on the LocationInsensitive variant: placeholder loans are very natural in this variant, as we don't compute subsets at all
• by adding the subset errors to the LocationInsensitive variant, we can re-enable it as a pre-pass in the Hybrid variant: if no potential illegal access errors nor potential illegal subset relations errors are detected, there's no need to run the full analysis.

I've also started to number and document the rules, as preparation for putting them in the book (and extracting them as soufflé .dl files if need be). This can be seen in the Naive variant (commit 321c23e, before the computation was simplified), and if that's interesting I'll apply it to the other variants (though the Opt rules are already more documented than the 2 others).

As the last bit of renaming, it also aligns the relation name and input fact files of the known_subsets to what the finalized rules look like in our hackmd (now known_placeholder_subsets). The commits starting with "tests" are the ones renaming input files and can be glossed over/ignored.

[lqd]

Working on documenting the LocationInsensitive variant in #126 made me realize I forgot to rename two relations in this PR here:

• requires was renamed to origin_contains_loan_on_entry in the rules hackmd, but this name is specific to the Naive variant: this variant here does not track locations, and should be origin_contains_loan
• known_contains still uses its old name dual to known_subset while the rules have it as placeholder_known_to_contain

I'll take care of both of these later, just to make my life easier until we have decided how to tackle #157, but I've used the correct names in the documentation PR.

[nikomatsakis]

r=me with nit