[WIP] Optimize bash and python

[svanoort]

Replaces #72 since similar changes have gotten pulled into
other PRs and I don't want merge conflicts.

Benchmark: time cat data/huwikisource-latest-pages-meta-current.xml | mycommand > /dev/null
System: Macbook Pro mid-2014 edition, 2.2 GHz Core i7
Times: real / CPU/ sys

Bash optimization history:
0. Baseline: 30.472s / 34.201s / 0.953s - output md5sum: 8ed673317ddae1eca0f0f76bb4b4605e

1. Bash script with more efficient sed + prefiltering + memory buffering of sort: 29.718s / 32.341s / 0.611s - output md5sum: e626474108822357e59507b26fa90f6a

Need to reconcile that difference here, it was not that way before

Note that with Mac we don't get parallel sort here. If I run the second one in an ubuntu:16.04 docker machine on the same host.

Bash TODOS:

• Verify bash generates correct output on Mac vs. Linux (line separator & UNIX vs Linux differences are causing problems)

Python optimizations:

1. Baseline: 25.631s / 25.145s / 0.493s - output md5sum b52070c4fdd3b0b87ea5659e2e862128
2. Remove UTF-8 encode/decode, and directly sort list: 14.633s / 14.302s / 0.368s - output md5sum: b52070c4fdd3b0b87ea5659e2e862128
3. Group together words with same count for sorting: 11.633s / 11.367s / 0.315s - output md5sum: b52070c4fdd3b0b87ea5659e2e862128
4. Rather than trying to sort count,index integers, collect lists of words with equal count directly 9.432s / 9.346s / 0.233s: - output md5sum: b52070c4fdd3b0b87ea5659e2e862128
5. Radix sort (or lexicographic sort using a lookup table, which is mapped onto a radix sort).