Is your feature request related to a problem? Please describe.
Currently, acceleration structure building does not handle vertex positions with Nan in them. Building an BLAS with a buffer that has a vertex containing a Nan is UB in both DX12 and Vulkan (in theory the current code can work with Nans in the x position due to some interesting mechanics).

Describe solutions you've considered

1. Copy vertex buffer into a temporary buffer
   We could copy every vertex into a new buffer, zeroing Nans as we went (e.g in a compute shader)
   • Pros
     • Simple
   • Cons
     • Requires a lot of memory for large scenes (as we have to keep all these buffers around until the command encoder ends)
   • Additionally, instead of disposing of these buffers, they could be placed into a vec to be reused. BLAS build could also be split if the combined buffer uses too much space.
2. Remove Nans from users buffer
   We could state that building BLASes may modify memory if there are Nans in the bit representation
   • Pros
     • Simple
     • Low memory footprint
   • Cons
     • Really confusing for users - No other API I know of can modify any buffer except the scratch buffer
3. Change Nans out of being Nans and then back to being Nans
   Use the fact that a Nan is a fully set exponent - Store a packed bool array (a bit per vertex component) whether a float is a Nan and if it is a Nan, remove a bit from the exponent (re-adding it after building).
   • Pros
     • Low memory footprint
   • Cons
     • Complex

I think that option 2 is probably not good, its confusing for users who would get their buffers changed under very edge case scenarios. Option 1 vs option 3 probably depends on how large real world vertex buffer usage is.

Additional context
I can't find anything in the Metal docs on whether Nans are allowed or not, so it might be UB.