Fix read_pixels_of_type UB: use bytemuck::cast_slice

[kimjune01]

Summary

• Replaces unsafe Vec::from_raw_parts + mem::forget with safe bytemuck::cast_slice + to_vec()
• The previous implementation had two sources of UB:
  1. mem::forget received a Vec<u8> whose memory was already transferred to another Vec
  2. Vec<T> would deallocate with align_of::<T>() but the memory was allocated with align_of::<u8>() = 1
• The new implementation safely casts the byte slice and copies into a properly-aligned Vec<T>
• bytemuck is already a dependency of fyrox-graphics

Test plan

• cargo check -p fyrox-graphics passes
• CI should verify no regressions across the full test suite

Closes #827

[mrDIMAS]

.to_vec call is kinda concerning here, I'm pretty sure that this can be done without calling it. Probably by creating a proxy struct that keeps the memory layout like in here - https://github.com/FyroxEngine/Fyrox/blob/master/fyrox-impl/src/scene/mesh/buffer.rs#L269-L277

[kimjune01]

@mrDIMAS thanks for the pointer to BytesStorage — addressed in 1d26960.

Introduced a TypedPixels<T: Pod> proxy struct that owns the original Vec<u8> and exposes &[T] via Deref (using bytemuck::cast_slice). The byte buffer is kept intact, so it's freed with the same Layout::array::<u8>(cap) it was allocated with. No .to_vec(), no copy.

I didn't carry an explicit Layout field like BytesStorage does, because that field is only needed when BytesStorage later decomposes back via from_raw_parts and deallocates manually. Here the Vec<u8> is never decomposed — dropping it as Vec<u8> already uses the right layout. Happy to add the field for symmetry if you'd prefer.

The single caller (fyrox-impl/src/renderer/hdr/mod.rs:207, evaluator.average_luminance(&pixels)) compiles unchanged via deref coercion. cargo check -p fyrox-graphics and cargo check -p fyrox-impl both pass.

[du82]

This "Kim June" guy is using AI to generate and submit pull requests. I suggest blocking him.

[mrDIMAS]

Yeah, but these changes were super obvious and I'm kinda ok with that. Anything more complex might not pass the review, because it would require real knowledge :)