changed to using decimal for conversion

[fcollman]

Here's a suggestion for how to resolve #719, not sure if i properly anticipated all the problems this might cause.

[fcollman]

@jbms any comments on whether this approach seems ok?

[jbms]

On the one hand, it seems always better to avoid losing precision when possible.

However, Neuroglancer itself uses floating point operations and therefore I think it is just chance that it happened to work here. In particular, while Neuroglancer happens to get the scale of the z dimension as 0.00003m exactly, it ends up with a scale of 6.500000000000001e-7 for x and y.

I created a separate fix of the underlying "instability" in Neuroglancer: #742

With that PR, things work as expected even with the mismatch.

There might still be some room to improve things on the Python side to make the behavior more consistent with Neuroglancer. But I'm not sure that using infinite precision to do the calculations is the answer.