[rlsw] Review depth formats and fix depth writing

[Bigfoot71]

I reviewed the depth formats and removed the 8-bit one. It could have been marginally useful for orthographic projection, but without heavier adjustments it would never really be practical.

I also added a 32-bit floating-point depth format, which avoids conversions, though after testing it turned out to be slightly slower than 16-bit with conversion. So 16-bit remains the default.

I also added saturation when writing depth values, since in some cases, when the camera was very close to a clipped triangle, the depth value could exceed the limit and cause an overflow.

This might require reviewing some details in the clipping, I find it strange at first glance...

This issue seems to mention the related UB problem: #5309

I compiled in debug mode with UBSan and ran several examples, didn't notice any issues on my side with this update.

The 24-bit format has also been quickly reviewed and tested, everything seems perfect!

[Bigfoot71]

After some more testing, the results are mixed about 32-bit format.

I found it strange that the 32-bit format was slower, but it seems to vary depending on the example, sometimes it's very slightly slower or basically identical, but in the first_person_maze example it actually gains just under 10 FPS.

Well, it would probably need more specific profiling.

[raysan5]

@Bigfoot71 thanks for the review! Definitely the 8-bit depth buffer use is very limited.