Replace HashMaps with a bit-vector for unique depth computation

[sampsyo]

Just a low-hanging fruit optimization for this op: instead of a Vec<HashMap<usize>>, we can use a single BitVec.

I did this in three steps and measured their performance on the chr22 GFA:

• Switching to a dense vector, but not a proper bit vector: 14.7 -> 2.2 seconds
• Switching to a proper bit vector from the bit-vec crate: 2.2 -> 2.0 seconds
• Reusing a single bit vector instead of allocating in the loop: no effect; maybe the optimizer already got this

Anyway, a quick profiling run suggests that running fgfa depth > /dev/null on this graph is now 90% printing stuff and 10% actual depth computation, so I think we'll call that good.