Make empty output type stable and have correct empty shape

[IanButterworth]

The output when no detections were made was unnecessarily type unstable, returning a CUDA array if GPU enabled, and for no real reason resolved to a different empty array size.
Given that the end filtering always happens on cpu, this can be simplified to always return Matrix{Float32} with the correct dim 1 size based on the model class count + standard params.

On a CPU-only machine

Master

No detections

julia> yolomod(batch, detectThresh=0.5, overlapThresh=0.8)
Any[]

1 detection

julia> yolomod(batch, detectThresh=0.5, overlapThresh=0.8)
89×1 Matrix{Float32}

This PR

No detections

julia> yolomod(batch, detectThresh=0.5, overlapThresh=0.8)
89×0 Matrix{Float32}

1 detection

julia> yolomod(batch, detectThresh=0.5, overlapThresh=0.8)
89×1 Matrix{Float32}