Distinguish scalars and tensors

[ev-br]

Over at #51 (comment), @lezcano observes that

we don't want to wrap scalars into tensors if we don't need to. Scalars are easier to treat in a compiler than tensors, and often get extra optimisations (no memory allocations, no synchronisations when accessed if we are working on CUDA, you can specialise on them...)

Note that this applies in general, so we should be mindful about this going all across.

So far the approach is quite the opposite: convert everything so scalars become 0-dim tensors, and proceed with tensors. Special-casing scalars would make wrappers heavier so if this is something we are going to do, it needs to be done across the codebase indeed.

[ev-br]

This is now at least consistent: we implement NEP51 in (a whitelisted subset of) binary ufuncs, the rest is probably low-prio.