Parser Improvements and newly labeled speeches

[jutton1]

Sorry about the delay with the PR... was busier than intended.

I also spent around 30 min a day hand labeling starts of speeches within sections of the record.

Both the old and new parses were run over the whole corpus and compared over 858,732 documents

	count	share
documents unchanged	854,079	99.46%
documents with ≥1 changed speech item	4,653	0.54%
speech items unchanged	2,682,217	99.665%
speech items newly attributed	5,096	0.189%
speech items reattributed	1,986	0.074%
speech items, boundary moved	1,703	0.063%
speech items no longer attributed	220	0.008%
words unchanged	576,175,058	99.755%
words newly attributed	567,175	0.098%
words reattributed	834,621	0.144%
words no longer attributed	15,897	0.003%

29.7% of speech items (799,542) get their speaker from the fallback rather than from
mods.xml, but for proof of concept, I reparsed without as mods.xml was hiding some of the failings of the old parser.

	with mods.xml	without	ratio
documents changed	4,653	63,392	13.6×
speech items newly attributed	5,096	59,030	11.6×
words newly attributed	567,175	15,719,712	27.7×
words no longer attributed	15,897	16,671	1.05×

The hand auditing was 250 windows of 40 lines each. 150 were centred on an item the
change moves, 100 from documents that parse identically. I did do this blind, so i had no clue which was which.

Metadata shows I spent around 182 minutes on hand labeling. I have placed the hand labeled in tests, I'll wire in something to utilize it later.

	exact-match	95% CI	precision	speaker IoU
old	70.4%	[66.8%, 73.8%]	98.7%	69.9%
this PR	99.7%	[98.9%, 99.9%]	99.5%	99.4%

Exact match means the rate at which the old and new parser matched what I hand labeled.
Precision is the rate at which i matched what the parser stated.
Speaker IoU is exact match over union of parser + human

closes #36 , closes #77, closes #79 , closes #80 , closes #81 , closes #82 , closes #83 , closes #84 , closes #85 , closes #86 , closes #87 , closes #88 , closes #89