Support constructing FMBScene with Python data

[horizon-blue]

Part of MET-36

Summary of Changes

While working on the forward kernel, I noticed that currently we don't have a way to create a FMBScene with data from Python side. This PR fills in this missing piece by defining a new constructor for FMBScene that takes in references to std::vectors and copy the values.

Sadly, this copy is needed because FMBScene is designed to manage the memory for itself and we want to avoid double-free issue. In addition, since Python list is always on host (CPU) side, if we want the scene to live in GPU memory, then we have to move the data anyways.

Aside: Using RAII Containers in FMBScene

While writing the constructor & the unit test, I started to get a bit frustrated by having to write similar code twice to specialize for host/device memory management, so I replaced the underlying containers in FMBScene with thrust::device/host_vector :). Another nice side effect of this is that Thrust's device pointer & device reference types handle the moving of data for us automatically when we attempt to read it from host (see the definition of FMBScene.__getitem__ in bindints.cu for example).

Test Plans

I added a unit test to verify that we can create FMBScene on both device & host and read the values stored in it.

pixi run test

[linear]

MET-36 Integrate and test the forward pass

Definition of done:

• the results of render_fmbs function calls should agree with the JAX code up to numerics
  • generate a teddy bear image
• [ ]