These techniques (eg von neumann) all seem to suffer from that same problem, namely, they crank out uniformly distributed random bits from biased sources, but with no guarantee of how long you may have to wait for a bit to come out.
Now, suppose that you had a function that turned N input bits with a 1:2 bias into a single bit with no bias. Then you could use this to implement an (0,1,2)-function as above, by collapsing inputs 1 and 2 into input 1. Since such an (0,1,2)-function is impossible, the 1:2-biased function is similarly impossible.
This argument holds for any i.i.d. input source that can be generated from a uniform-integer source with an odd number of possible values.
And, as a matter of fact, much faster than the input if required.
This can be mitigated if you use a unkeyed hash function instead, but even for these we don't have any provable guarantees unless you assume the function is a random oracle, which is an ideal object that cannot exist.
https://en.wikipedia.org/wiki/Fortuna_(PRNG)
[EDIT]: Also: https://words.filippo.io/dispatches/linux-csprng/