Update benchmark sampling method and statistics

[popematt]

Issue #, if available:

While not necessarily a fix, this does make significant progress on #265

Description of changes:

• Updates/fixes the sampling methodology so that it's properly running samples of 1-2 seconds each.
• Adds a SampleDist class that helps compute statistics for a sample distribution
• Drops the "rate" metrics and adds operations per second metrics.
• Drops the percentile stats and adds error and standard deviation.
• Refactors the report generator to make it less spaghetti-like
• Updates the table output format so that it can be directly copy-pasted to github as a markdown table.
• Updates the compare command so that the fields argument uses the same field names and syntax as the other commands.
• Drops support for Python 3.7 (which is EOL) because the statistics.NormalDist class requires at least Python 3.8.

The perf testing workflow is expected to fail as I have not updated that. It will come in the next PR.

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license.

[popematt]

Using the new table format, you can copy/paste directly from the terminal to e.g. a github comment, and this is what you will see:

name	time_mean(ns)	time_error(ns)	ops/s_mean	ops/s_error
(cbor,dumps,cat.cbor)	621.56	6.53	1609325.62	16351.13
(cbor2,dumps,cat.cbor)	2559.77	20.53	390726.48	3086.54
(json,dumps,cat.json)	1619.62	6.14	617453.26	2341.06
(ujson,dumps,cat.json)	698.16	3.01	1432404.24	6223.26
(ion_text,dumps,cat.ion)	9420.41	84.96	106175.15	938.70
(ion_binary,dumps,cat.ion)	10463.80	59.24	95575.76	541.19
Ion Binary, pure Python impl	263795.71	1908.30	3791.34	27.22
Ion Text, pure Python impl	107178.88	1369.13	9334.16	115.81
(self_describing_protobuf,dumps,cat.sd_protobuf_data)	3974.29	55.22	251742.58	3371.29
(protobuf,dumps,cat.protobuf_data)	285.56	2.89	3502871.19	34420.39

Note that this suffers from false precision (just like JMH reports do). The number formatting needs to be fixed to show a specific number of significant figures—possibly dependent on the error value for the results.

[tgregg]

Nice. The autorange feature was a good find.