drop support for python 2

The common unit is microseconds. While this value was correct, its name
was misleading.

Otherwise, if `url` contains non-ascii chars, startswith will fail with
an error like:
  UnicodeDecodeError: 'ascii' codec can't decode byte 0xc3 in position 84: ordinal not in range(128)
becasue 'http' is unicode. So byte-ify it to avoid this transcoding.

Give them captions.
Also, tread 499 as a non-error: client closed the connection before server
could respond, so it is likely not something the server could be
considered responsible for. Of course, the response time still matters,
so if these statuses come after slow responses it will still affect the
score.

Since 2.4.13, httpd suports %{UNIT}T in LogFormat

Traceback (most recent call last):
  File "apachedex/__init__.py", line 1274, in wrapper
    return func(*args, **kw)
  File "apachedex/__init__.py", line 1586, in main
    site_data.rescale(rescale, getDuration)
  File "apachedex/__init__.py", line 605, in rescale
    for value_date, data in getattr(self, attribute_id).iteritems():
NameError: global name 'attribute_id' is not defined

Allow to define how many pages will be displayed in 'Hottest pages' section.

By default (without passing this option), there is a single row grouping
results of all non-modules URLs ('other'). With this new options, one row per
URL is displayed for non-modules results.

Unlike apache which escape non ascii characters in referrer, caddy
writes referrer as is. Edge seem to send referrer not escaped, so with
Edge and caddy we can have non ascii text in referrer.

For lines which cannot be decoded as ASCII, we use python `replace`
error handler which would in this case allow the line to be processed if
the decoding problem is only about the encoding of the referrer.

We don't implement this case as "skip and report ill-formed line",
because python does not provide utilities to do this easily.

Reproduction with caddy:

```
curl -k http://localhost -H 'Referer: héhé'
```

With apache, `LogFormat "%h %l %u %t \"%r\" %>s %b \"%{Referer}i\" \"%{User-agent}i\"" common`
```
127.0.0.1 - - [28/Feb/2019:10:03:33 +0100] "GET / HTTP/1.1" 200 2046 "h\xc3\xa9h\xc3\xa9" "curl/7.50.1" 4
```

With caddy, `log / stdout "{remote} {>REMOTE_USER} [{when}] \"{method} {uri} {proto}\" {status} {size} \"{>Referer}\" \"{>User-Agent}\" {latency_ms}"`

```
127.0.0.1 - [28/Feb/2019:10:05:00 +0100] "GET / HTTP/2.0" 200 1950 "héhé" "curl/7.50.1" 4
```

When the table is large and requires scrolling, it becomes difficult
to read so add 'title' attribute to <tr> on 'Stats per module' and
'Hits per status code' tables. For example when hovering a particular
cell on a module row, it display the module name.

/reviewed-on nexedi/apachedex!3

We observed lines in our logs where the timestamp field was still
respecting the timestamp regexp, so the line was not reported as
invalid, but parsing such timestamp caused a ValueError in
_matchToDateTime

The beginning of line was:
127.0.0.1 - - [14/Jul/2017:127.0.0.1 - - [14/Jul/2017:09:41:41 +0200]

Which uses `[14/Jul/2017:127.0.0.1 - - [14/Jul/2017:09:41:41 +0200]` as
timestamp, so this fail the simple .split() used to separate timestamp
and timezone.

Added a minimal test case to reproduce this specific problem.

Use a decorator to not indent the whole function.

This change introduces a (minor) parameter value backward-incompatibility.

There is no standard unquoted URL encoding (=charset), so do not unquote
to just handle ASCII. Fixes various UnicodeDecodeError failures, at the
cost of decreased report readability.

This causes errors and solves nothing.

Do it while parsing, so (for example) day splitting happens at midnight in
chosen timezone instead of midnight in source timezone.
When no timezone is provided, use input's times verbatim.
DST changes cause artefacts in graphs (double or zeroed hour).
Also, display used timezone in generated page.

Depending on input, some state files can starts with ".", and are not
matched by "state/*".