Add NotImplemented and NotImplementedRule

[ararslan]

Sometimes you just don't implement something. Maybe it's hard, maybe you're lazy like me, whatever. For such cases, there is the differential NotImplemented and its associated rule NotImplementedRule.

This was born of not being able to figure out the rule for the scalar multiple parameter in BLAS.gemm (#25), but having implemented rules for the matrix parameters. As indicated by the previous paragraph, I am lazy; it is easier to add a rule that admits that to the world than it is to actually figure out how to differentiate wrt alpha.

[oxinabox]

Can there be a "not yet" added here?
Maybe even a "PRs to remove not implemented rules are welcome"

[ararslan]

Good idea.

[ararslan]

The more I think about this, the less I like it.

[nickrobinson251]

Seems fine. Seems like it might be useful. Shall we do this?

[ararslan]

Up to you/whoever maintains this now. I won't likely have time to revisit this but feel free to do what you will with the branch.

[oxinabox]

Yes, I want to return this instead of Nothing as the default fallback

[willtebbutt]

As with @oxinabox 's comment above, I would like to see this expanded slightly to say more about what a NotImplementedRule does mean. Some context would be helpful e.g.

"""
Rule indicating that a particular derivative is not yet implemented by ChainRules. Different from [`DNERule`](@ref) in that this rule does not imply non-differentiability, but rather that no one has implemented it yet.
"""

[willtebbutt]

I'm very happy for this to be merged provided that someone addresses the docstrings to provide a bit more context for the confused rule-implementer.

[nickrobinson251]

Cool, I can take a look at finishing this up next week, unless someone else is interested in doing it :)

[oxinabox]

Yes, I want to return this instead of Nothing as the default fallback

For context currently we do:
https://github.com/JuliaDiff/ChainRulesCore.jl/blob/ccead4ffa606cd9e83f5551e3767dac212ebddaa/src/rules.jl#L352
rrule(::Any, ::Vararg{Any}; kwargs...) = nothing

The AD system when it gets that back, know that it has to actually go and do AD.

I was proposing we change that to:
rrule(f::Any, args::Vararg{Any}; kwargs...) = (f(args...; kwargs...), NotImplementedRule())

Advantage of this is internal consistancy, and less magic Nothing floating around.
Disadvantage is we would be calculating the forward result (f(args...; kwargs...)),
and the AD system is going to want to calculate the forward result itself also when finding the sensitivity/gradient.
Which is pretty bad -- makes it take twice as long. So lets not do that.

This case does still show up for case where 1 of the derivatives is listed as NotImplementedRule() though.
Which would require the same thing.
However, hopefully that is rare because the NotImplementedRule would be used rarely,
and only for things where it was unlikely to matter.

[oxinabox]

I had a discussion with @willtebbutt about this.
In particular, that last point about doing the forward pass twice.

And we concluded that for now we don't need it,
and it complicates things and causes problems,
so lets not do it.

I am closing this PR, but we can reopen it later when/if we need it.