> And making float('nan') returning a singleton, 
> but 1e1000 * 0 returning different NaN would cause large confusion.

Not really, it would be just be an implementation detail, no different than int and strings being sometimes unique and sometimes not.  Historically, immutable objects are allowed to be reused when it is convenient for the implementation.

> What about Decimal NaN?

Decimal isn't used as much, so the need is less pressing, but we can do whatever is allowed by the spec.  Presumably, that would be easier than with floats because we control all possible ways to construct Decimals.