Change handling of OpenBlas buffers

[oscardssmith]

Currently, most of the ram used by julia on startup is buffers for OpenBlas. This is fine for numerical code, but wastes a bunch of memory for code that doesn't multiply matrices (I was as shocked as you are to discover that not everyone multiplies matrices). I'm not sure what the best way to achieve this is, but in Python, loading NumPy doesn't allocate a ton of ram. Given this, deferring the allocation of OpenBlas buffers until matrices are actually multiplied would increase the responsiveness of Julia for non-numeric use-cases.

[ViralBShah]

I'm making this a discussion, since there isn't a clear plan on what needs to be done here.

[PallHaraldsson]

Is it a problem? I mean is the memory allocated and then not used? Then doesn't cost physical RAM? However LinearAlgebra is being excised, and with it OpenBLAS I guess, I'm at least hoping that. BLIS.jl better anyway(?!).