Improvements to PCSS shadow sampling

[mvaligursky]

Related to PCSS implementation: #5483

• consider what needs to be done with the hardcoded constant here: Address PCSS softness #5483 (comment) DONE Overhaul of directional PCSS shadows #7258
• the implementation uses hardcoded 16 samples #define PCSS_SAMPLE_COUNT 16 - perhaps add some control over this DONE Overhaul of directional PCSS shadows #7258
• Make compatible with WebGPU, at least for directional lights. WebGPU only supports clustered lights currently, so other lights won't work unless PCSS is supported by clustered lights. DONE Overhaul of directional PCSS shadows #7258
• R32F texture is used to store shadow depth map. We need a fallback for platforms that do not support it, perhaps the same as RenderPassPrepass uses, which stores the depth float bits in RGBA8, which then SSAO decodes back to float. DONE - added F32->F16->PCF3 fallback here: 3b1e69f
• adjust omni and spot PCSS code to work as well as directional updated in Overhaul of directional PCSS shadows #7258
• add sample randomization to allow smaller number of samples then TAA is enabled
• We currently render to F32 texture + depth buffer, but only use the F32 texture. For directional lights (and eventually for all types) we store non-linear depth, meaning the color buffer contains the same data as depth buffer. On WebGPU, we should be able to avoid using color buffer entirely, and simply use depth buffer only - as long as we can load a single sample from it.

[mvaligursky]

updated in #7258