Strip [] from IPv6 addresses for FastCGI.

Fixes #845.

systemd, README: Edit to account for the recent spike in reports

We have had three operators within a few days which ran into the same cause
and had not been able to figure out what went wrong.

addresses #833, #822

fix typo

Provides some more guidelines to operators on how to avoid running Caddy as root.

Introduces an user www-data, which really is a placeholder. Such an user with the same UID/GID combination is created on the most popular Linux distribution. I trust any operator can spot the difference to his/her distro and adjust the unit file.

User nobody is not used here to avoid two easy pitfalls: Such an user should not be able to access private keys (for TLS), and should not write private keys (we would do that with Letsencrypt).

The exemplary unit file for systemd is intentionally redundant at times, for
example dropping privileges which an unprivileged user "www-data" did not have
in the first place: To aid as fallback in case the file gets copied and an
operator setting UID to 0 (which reportedly happened in the past).

browse: Decorate external links with: noopener noreferrer

Setting these on external links prefents the target from manipulating this
page by "window.opener" with some widely deployed browsers.

Restart gracefully with in-process restart

Adding note on the rationale for Go 1.6

Markdown changes

* Overwrite proxy headers based on directive

Headers of the request sent by the proxy upstream can now be modified in
the following way:

Prefix header with `+`: Header will be added if it doesn't exist
otherwise, the values will be merge
Prefix header with `-': Header will be removed
No prefix: Header will be replaced with given value

* Add missing formating directive reported by go vet

* Overwrite up/down stream proxy headers

Add Up/DownStreamHeaders to UpstreamHost

Split `proxy_header` option in `proxy` directive into `header_upstream`
and `header_downstream`. By splitting into two, it makes it clear in
what direction the given headers must be applied.

`proxy_header` can still be used (to maintain backward compatability)
but its assumed to be `header_upstream`

Response headers received by the reverse proxy from the upstream host
are updated according the `header_downstream` rules.
The update occurs through a func given to the reverse proxy, which is
applied once a response is received.

Headers (for upstream and downstream) can now be modified in
the following way:

Prefix header with `+`: Header will be added if it doesn't exist
otherwise, the values will be merge
Prefix header with `-': Header will be removed
No prefix: Header will be replaced with given value

Updated branch with changes from master

* minor refactor to make intent clearer

* Make Up/Down stream headers naming consistent

* Fix error descriptions to be more clear

* Fix lint issue

* Move handling of headers around to prevent memory use spikes

While debugging #782, I noticed that using http2 and max_fails=0,
X-Forwarded-For grew infinitely when an upstream request failed after
refreshing the test page. This change ensures that headers are only
set once per request rather than appending in a time-terminated loop.

* Refactor some code into its own function

Another new safeguard is that we check whether the datetime has been read
correctly. If not then the listing will not be localized.

closes #793

[1] https://github.com/golang/go/blob/57e459e02b4b01567f92542f92cd9afde209e193/src/crypto/tls/common.go#L424
[2] https://github.com/golang/go/blob/57e459e02b4b01567f92542f92cd9afde209e193/src/crypto/tls/common.go#L392-L407

[2] has overwritten the first tls ticket key on round N=0, that has previously
been written using [1].

Go's stdlib does not use c.sessionTicketKeys≥1 as indicator if those values had
already been set; initializing that lone SessionTicketKey does the job for for
now.
    If c.serverInit() were called in round N+1 all existing tls ticket keys
would be overwritten (in round N<4 except the very first one, of course).
As member variables of tls.Config are read-only by then, we cannot keep
updating SessionTicketKey as well.

This has been escalated to Go's authors with golang/go#15421 here:
https://github.com/golang/go/issues/15421

Thanks to Matthew Holt for the initial report!