Allow switching devices. Use `__init__`.

[lucasb-eyer]

This introduces the use_current_device function which initializes CUBLAS for the currently active device, thus allowing the use of blas.jl functions on non-default GPU.

In addition, move initialization (and registration of the finalizer) into __init__ so that we can make use of pre-compilation in the future.

The default behavior should be unchanged by this.

[kshyatt]

Hi @lucasb-eyer, I'm sorry it took so long for someone to look at this. This PR LGTM and I'll merge it tonight if no one has any objections.

[lucasb-eyer]

No worries, I'm not in a hurry.

[lucasb-eyer]

ping, I guess =)