Fixes #3701.

Thanks @ExcaliburZero for noticing. Fixes #3698.

* Sky: rename Box => m_box and inline getBoundingBox
* Uniformize aabbox3d<f32> to aabb3f

This makes the functions a bit faster since they don't
have to recreate the tables every invocation, and makes
the code more readable.

Also, document `wallmounted_to_dir`.

The function was implemented but not documented in `lua_api.txt`.

Introduce a new contentfeatures version (8). When clients
connect using v27 protocol version, they can assume that
the tiledef.backface_culling is trustable, but if clients
connect to servers providing contentfeatures version 7,
then the v27 clients know that backface culling settings
provided by the server in tiledefs are bogus for mesh,
plantlike, firelike or liquid drawtype nodes.

thanks to hmmmm, est31, nerzhul.

Tested on new client - new server, new client - old server
old client - new server.

Also use string::empty method, it is better than comparing with empty strings.

All places that lock this mutex are only called by the main thread:
ClientMap::updateCamera(), ClientMap::updateDrawList(), ClientMap::renderMap(), ClientMap::renderPostFx().

To avoid spawn search failing in new specialised mapgens
Increase spawn search range to 4000 nodes
Add getSpawnLevelAtPoint() functions to EmergeManager, class Mapgen
and all mapgens
Remove getGroundLevelAtPoint() functions from all mapgens except mgv6
(possibly to be re-added later in the correct form to return actual
ground level)
Make mgvalleys flag names consistent with other mapgens
Remove now unused 'vertical spawn range' setting

The engine implementation of `print` packs the varargs into a
table and passes the table directly to `table.concat`. If you
pass any value not supported by `table.concat` (particularly
`nil`), the server crashes. This is unexpected behavior, as
`print` is supposed to be able to work with anything.

This patch changes the implementation so it first converts
all arguments using `tostring`, which fixes the issue and
makes the custom `print` function compatible with the stock
Lua behavior.

-> Replace the three bool params with an enum
-> Add struct for the static content, leads to less repetition
-> cache enable_animations setting

Stop riverbeds from forming plateaus under sea. Minor corrections to
random lava/water placement.

Tunnel 3D noises are only calculated when solid terrain is present
in mapchunk, avoiding large amounts of unnecessary calculations
Change 'int' to 's16' in calculateNoise
Change 'i' to 'vi' for voxelmanip indexes for consistency
Keep 'u32 index3d' local to a smaller part of tunnel code
Mgv7: Don't call CaveV7 if no solid terrain in mapchunk
Give 'open' bool a more descriptive name

Replace simple caves with V5 caves, adding unpredictable water and lava
settings and massive caves based on subterrain. Remove fast terrain mode
and accompanying settings. Remove superfluous temperature/humidity
settings. Remove lava/water height setting. Fix errors in humidity
handling and remove humidity_break_point setting. Move cave noises to
generateCaves. Fix minor formatting/naming issues and use
MYMAX/MYMIN/myround.

Non const references cause a lot of confusion with behaviour of code,
and are disallowed by minetest style guide.

I could honestly not make much sense of the timer implementation
that was here. Instead I've implemented the type of timer algorithm
that I've used before, and tested it instead.

The concept is extremely simple: all timers are put in an ordered
list. We check every server tick if any of the timers have
elapsed, and execute the function associated with this timer.

We know that many timers by themselves cause new timers to be
added to this list, so we iterate *backwards* over the timer
list. This means that new timers being added while timers are
being executed, can never be executed in the same function pass,
as they are always appended to the table *after* the end of
the table, which we will never reach in the current pass over
all the table elements.

We switch time keeping to minetest.get_us_time(). dtime is
likely unreliable and we have our own high-res timer that we
can fix if it is indeed broken. This removes the need to do
any sort of time keeping.

clock_gettime() is a far better clock than gettimeofday().

Even better than clock_gettime() is that you can select either
CLOCK_MONOTONIC, or even CLOCK_MONOTONIC_RAW. These clocks offer
high precision time. And the _RAW variant will never roll back
due to NTP drift or daylight savings, or otherwise.

I've adjusted this code to select the right clock method auto-
matically based on what's available in the OS. This means that
if you're running a very old linux version, MacOS or other,
you will automatically get the best clocksource available.

I've tested all Linux clocksources by selectively compiling and
running a 10k+ timer test suite. In all cases I confirmed that
the 3 POSIX Linux clocksources worked properly, and were
selected properly.

I've modified the OS X compile path to use the high-res clock
source for all time functions, but I can't confirm it works or
that it compiles.

As for WIN32, I confirmed that the used clocksource is indeed
a Monotonic clocksource, so good news: that code section appears
to be exactly what it should be.

`errorstream` must not be overly verbose as clientside it is directly printed
onto the ingame chat window. These days, the serverlist can contain > 200k bytes,
so better print it to warningstream if the data buffer is too long.

Currently translated at 52.0% (410 of 787 strings)

Currently translated at 32.4% (255 of 787 strings)

Currently translated at 95.5% (752 of 787 strings)

Currently translated at 41.6% (328 of 787 strings)

Currently translated at 26.6% (210 of 787 strings)