Actually compile with proper optimizations in a release build.

[BMagnu]

As it turns out, we did not actually compile the release builds with the maximum safe optimization level.
First of all, on Linux / Mac builds, we did not compile with -O3, but only with -O2. That's wasted performance, as O3 is still safe (unlike -Ofast, which may do inaccurate math).
Second of all, none of the toolchains has had Link-Time Optimizations enabled. This is especially bad in the FSO case, as we have some functions in cpp's (and thus uninlineable unless LTO is available) which are called very often (such as the vector math functions).
Enabling LTO throws up a few warnings: Some ODR violations, two uninitialized structs, and a few "potential negative size allocations". These latter ones are technically already caught with an Assertion, but as these are not compiled into release builds, the compiler sees a potentially dangerous path and complains. Until we get assumptions for the compiler (which is not soon, given that these are compiler extensions added to GCC13+ and standardized in C++23), we have to just make sure that the code at runtime ensures their correctness, even if that may be redundant.
Furthermore, an update to CMake 3.9 is required to actually get portable LTO support detection and enabling.
This also yields actually measureable performance increases.
In a particle stress test I was profiling (which prompted this PR), a performance boost of 10% is detectable (lower single graph / orange line is with LTO and -O3):

[BMagnu]

There seem to be some stray warnings left that my local GCC did not trigger, I'll need to fix them before this is merged / reviewed

[EatThePath]

Oh, nice find. Have you measured any performance change in more gamelike conditions than that stress test?

[BMagnu]

I want to do icarus at some point, but I didn't find the time yet. I'll probably get to it around the weekend, once I squashed the remaining warnings.
And, evidently, I can only speak for the linux performance, not for how much LTO actually does for MSVC here.

[BMagnu]

Okay, finally managed to benchmark Icarus. Upper Graph / Blue Line is master, Lower Graph / Orange is this PR.

As such, a small but consistent speedup can be observed.
Of course, this is quite compiler-dependent, so the actual speedup observed on MSVC may be different.
It is also to note, that the main bottleneck on my system is the very moody nvidia driver. As such, stuff like system / driver uptime has an effect on my benchmarks that far exceeds most changes (i.e. rebooting my system after a day of use usually nets +25% FPS with the nvidia driver). As such, take any benchmarks from me with some grain of salt when comparing it to others not directly measured at the same time.