To have debug logs dump before FLOG panic
and when it is necessary when debugging a rare issue.

In options_cleanup() according to valgrind not closing
logfd if it is the default stdout.

A few curl version after 8.9.0 kept connections opened
which did not count in CURLMOPT_MAX_TOTAL_CONNECTIONS limit.
Therefore more IO events are necessary to handle more
sockets than we expect.

Details in: curl/curl#15857

It was not working at all before.
Better to get rid of TCP keepalive since it was not necessary.
Raising curl requirement level because of option.
It has been made configurable via argument, because different
DoH server providers have different timeouts configured.

When a large burst of DNS requests came in and the
only HTTPS/2 connection was too many time in idle state,
curl closes it and opens brand new connection for every
single DNS request.
Obviously later it will keep only 1 and close all the others.
So it is better to wait for multiplexing.
When testing this setting, the DNS query latency values
did not raise.

If we already have a connection 5 second should be enough,
else leave time for connecting to server.
Goal is to fail faster if some reused HTTPS connection was idle
for way too many time and it is actually not working anymore.

It was quite a mess and inconsistent.
It will not be perfect with this change, but still better.
In case of boolean options, default values are omitted.

Fixes 14