rm UnsafeArrays, use previous way

[ericphanson]

this reverts the UnsafeArray related changes in #18 since they introduce memory safety issues. They are also not needed! I get equally good results with this PR for r3tex/ObjectDetector.jl#101:

forward pass: 0.907787 seconds (4.30 k allocations: 6.263 GiB, 5.95% gc time)
yolomod(batch, detectThresh = 0.5, overlapThresh = 0.8): 0.961214 seconds (47.01 k allocations: 6.298 GiB, 5.62% gc time, 2.79% compilation time)
forward pass: 1.293893 seconds (4.30 k allocations: 6.263 GiB, 21.28% gc time)
yolomod(batch, detectThresh = 0.5, overlapThresh = 0.8): 1.320383 seconds (6.89 k allocations: 6.295 GiB, 20.86% gc time)
forward pass: 0.924500 seconds (4.30 k allocations: 6.263 GiB, 4.77% gc time)
yolomod(batch, detectThresh = 0.5, overlapThresh = 0.8): 0.949379 seconds (6.89 k allocations: 6.295 GiB, 4.64% gc time)

forward pass: 1.132717 seconds (6.62 k allocations: 12.082 MiB)
bumper_yolomod(b, batch_aa; detectThresh = 0.5, overlapThresh = 0.8): 1.181193 seconds (26.75 k allocations: 13.084 MiB, 1.14% compilation time)
forward pass: 1.130855 seconds (6.62 k allocations: 12.082 MiB)
bumper_yolomod(b, batch_aa; detectThresh = 0.5, overlapThresh = 0.8): 1.167223 seconds (9.09 k allocations: 12.159 MiB)
forward pass: 0.787238 seconds (6.62 k allocations: 12.082 MiB)
bumper_yolomod(b, batch_aa; detectThresh = 0.5, overlapThresh = 0.8): 0.822033 seconds (9.09 k allocations: 12.159 MiB)

I think the vectorization thing was a red herring. Maybe the important bit is the no-op constructors to prevent nesting, which are retained here.

[ericphanson]

I wish I hadn't registered #18 but I did, so I guess I will make this another breaking release