Remove string cache and some traits

[SimonSapin]

This change is

[SimonSapin]

r? @nox

[SimonSapin]

At first I split the bloom filter into its own crate so that implementing BloomHash requires a smaller dependency, but instead I removed the trait and use std::hash::Hash with a FNV hasher instead.

My guess is that hashing will be slightly more expensive than before, but we might get fewer hash collisions. (The previous BloomHash for Atom would just XOR the two halves of the u64 inside Atom, whose bit pattern is very much not random.) I tried both in Servo with ./mach run -r https://en.wikipedia.org/wiki/Servomechanism -p -x and did not see a significant difference either way in "Style Recalc" times.

This PR also weakens a test that check the number of false positives in a bloom filter. I believe this test only passed before because it hashes consecutive integers, which the previous BloomHash for usize would hash to themselves with no collision at all. In practice the bloom filter is used with atoms, not consecutive integers.

[nox]

@bors-servo r+

[bors-servo]

📌 Commit 7346186 has been approved by nox

[bors-servo]

⌛ Testing commit 7346186 with merge e50ee6a...

[bors-servo]

☀️ Test successful - status-travis