I removed the implementation, and changed the tests to use `rgb32frame-crop.bmp` to avoid problematic cropping of the full-sized input. We don't need an implementation currently since there are no cases of this in the wild, and the D3D-aware case can't be tested anyway.
The issue can be fixed by passing an aligned width to `MFCalculateImageSize()` if NV12 is used, and fixing up the output buffer. This requires a locally allocated sample, for which it's probably possible to use the `MFT_OUTPUT_STREAM_PROVIDES_SAMPLES` pathway with little modification.
This is ready for further review now.