Experiment with smaller depth buffer formats

[JMS55]

Currently we're using 32 bit depth buffers.

We could probably get away with 24 bit depth, or maybe even 16 bit depth for shadow maps once we improve precision there.

Less bits = faster to write and read.

[IceSentry]

To give some more context, OIT already uses 24bit depth and IIRC so does deferred. So it probably makes sense to just do that everywhere.

[mrec]

Are you confident that 24-bit depth is natively supported on the target platforms you care about? I believe it's sometimes emulated using 32-bit, at which point you're just throwing away precision for nothing.

See e.g. https://gpuopen.com/learn/rdna-performance-guide/

24-bit depth formats have the same cost as 32-bit formats on GCN and RDNA. Use a 32-bit format on AMD to get higher precision for the same memory cost as 24-bit. 16-bit depth does not necessarily compress better than 32-bit depth.

Note that Godot switched to using 32-bit for desktop and 24-bit for mobile earlier this year: godotengine/godot#87057