* Proxy can now use QUIC for upstream connections

Add HandshakeTimeout, change h2quic syntax

* Add setup and upstream test

Test QUIC proxy with actual h2quic instance

Use different port fo QUIC test server

Add quic host to CI config

Added testdata to vendor

Revert "Added testdata to vendor"

This reverts commit 959512282deed8623168d090e5ca5e5a7933019c.

* Use local testdata

Renewed certificates would not be reloaded into the cache because their
names conflict with names of certificates already in the cache; this
was intentional when loading new certs to avoid confusion, but is
problematic when renewing, since the old certificate doesn't get
evicted from the cache. (Oops.)

Here, I remedy this situation by explicitly deleting the old cert from
the cache before adding the renewed one back in.

httpserver: Fix #1859 by cleaning paths when matching them

Signed-off-by: Matthew Holt <mholt@users.noreply.github.com>

Do not push index file when not in a rule

+ test

* Add tests for SCRIPT_NAME

* fastcgi: Include vhost path prefix in SCRIPT_NAME

* Allow pushing multiple resources via Link header

* Add nopush test case

* Extract Link header parsing to separate function

* Parser regexp-free

* Remove dead code, thx gometalinter

* Redundant condition - won't happen

* Reduce duplication

The race was in the test only; not in the production code

* browse: Attempt to fix tests on Windows

* browse: Make tests verbose for debugging

* Moar debugging

* Trying path.Join instead

* browse: Just skip the tests for now

* browse: Remove debug prints

build.go is (should be) cross-platform compatible.

Timestamps are now excluded from all builds on a clean commit,
in an effort to be byte-for-byte reproducible.

* templates: Execute template loaded by later middlewares

This is the beginning of an attempt to make the staticfiles file server
the only middleware that hits the disk and loads content. This may have
unknown implications. But the goal is to reduce duplication without
sacrificing performance. (We now call ServeContent here.)

This change loses about 15% of the req/sec of the old way of doing it,
but this way is arguably more correct since the file server is good at
serving static files; duplicating that logic in every middleware that
needs to hit the disk is not practical.

* httpserver: Introduce ResponseRecorder as per Tw's suggestions

It implements io.ReaderFrom and has some allocation-reducing
optimizations baked into it

* templates: Increase execution speed by ~10-15% after perf regression

By using httpserver.ResponseBuffer, we can reduce allocations and still
get what we want. It's a little tricky but it works so far.

tls: Fix OCSP stapling bug when certificate names overlap other certs

Double systemd LimitNPROC to allow more threads

Add forwardproxy to directives' list

Fix restart restart behaviour of the systemd service

The previous setting caused the service to hit a rate-limit when it was
restarted more than 5 times in 24h.
Editing the Caddyfile and restarting the service could also easily
trigger this rate limit.
One could argue that users could simply call `systemctl reset-failed
caddy` to reset the rate-limit counter, but this is counterintuitive
because most users won't know this command and are possibly unaware that
they had hit a rate-limit.

The service is now allowed to restart 10 times in 10 seconds before
hitting a rate limit.
This should be conservative enough to rate limit quickly failing
services and to allow users to edit and test their caddy configuration.

This closes #1718

Remove restart limit settings and use defaults

By default 5 restarts within 10 seconds are allowed without
encountering a restart limit hit, see  `man systemd.unit` for details.

Set Restart to on-abnormal

The table in https://www.freedesktop.org/software/systemd/man/systemd.service.html#Restart=
shows the conditions for which on-abnormal would restart the service.
It will *not* restart the service in the following cases:

- a non-zero exit status, e.g. an invalid Caddyfile
- a zero exit code (or those specified in SuccessExitStatus=) and a clean signal
    clean signals are SIGHUP, SIGINT, SIGTERM or SIGPIPE
    https://github.com/systemd/systemd/blob/3536f49e8fa281539798a7bc5004d73302f39673/src/basic/exit-status.c#L205

The service *will be restarted* in the following cases:

- a unclean signal, e.g. SIGKILL
- on start and watchdog timeout (we don't use those systemd service
constructs explicitly)

Fix for #1794: Fixes issues with IfMatcher and regular expressions.

browse: Support absolute and recursive directory symlinks

* SIGUSR2 triggers graceful binary upgrades (spawns new process)

* Move some functions around, hopefully fixing Windows build

* Clean up a couple file closes and add links to useful debugging thread

* Use two underscores in upgrade env var

To help ensure uniqueness / avoid possible collisions

https://caddy.community/t/random-ocsp-response-errors-for-random-clients/2473?u=matt

Certificates are keyed by name in the cache, optimized for fast lookups
during TLS handshakes using SNI. A more "correct" way that is truly a
1:1 would be to cache certificates by a hash of the leaf's DER bytes,
but this involves an extra index to maintain. So instead of that, we
simply choose to prevent overlap when keying certificates by server
name. This avoids the ambiguity when updating OCSP staples, for instance.

(allow no arguments of paths to protect)