Allow rand to sample from non-identical Sampleables over time

[tpgillam]

See original comment below. On discussion, we think that this is a good thing to do, even though it is breaking.

Original comment:
 
@willtebbutt I'm actually wondering if it would be good to make a breaking change here, whilst working on this section.

Currently, this version of rand takes:

• A node x, which is used only for alignment
• (maybe) a constant S, determining the sampleable
• (maybe) some dimensions d (either tuple or int varargs), which make the output value type be an array

I think it might be more future proof to unify x and S.

Pros: it means that in the future, one can e.g. have a node that ticks with distributions, so that the random value drawn at each knot can be drawn from that distribution.

Cons: It is breaking (although better now than later), and it's slightly more awkward to generate "trivial" random numbers.

e.g. old:

x = pulse(Hour(1))
y = rand(x)

new:

S = align(Float64, pulse(Hour(1)))  # i.e. a node that ticks with `Float64` every time that we wish to sample
y = rand(S)

Any thoughts?

Originally posted by @tpgillam in #67 (comment)