Here's a description of the Mazur swindle as I understand it (see https://math.stackexchange.com/a/3348047/172988):
You take the $x$-axis in $\mathbb{R}^3$ and tie into it $K_1$ in the intervals $[2n,2n+1]$ and $K_2$ into $[2n+1,2n+2]$ for $n\geq 0$ an integer. Then, since we assume $K_1\mathbin{\#} K_2$ is isotopic to the unknot, we can take intervals $I_n=[2n,2n+2]$ and $J_n=[2n+1,2n+3]$ and individually isotope $I_n$ or $J_n$ back to the $x$-axis for a fixed $n$. Then, the idea is that you can simultaneously do all the $I_n$ isotopies to get the $x$-axis itself; similarly you can simultaneously do all the $J_n$ isotopies to get just $K_1$ tied into $[0,1]$. One can then conclude $K_1$ is isotopic to the unknot. I am only sketching things out here, and there are various topological details to check. Also, this infinite knotted object is not a knot; perhaps one might call it a "longer knot." The key is that everything can be done with piecewise linear topology, which makes the usual 3-manifold arguments work out.
This version of the argument constructs an actual isotopy of $\mathbb{R}^3$, but it can be turned into diagrammatic manipulations by Reidemeister moves, in a certain sense. The only complication is that infinitely many non-conflicting Reidemeister moves are performed in parallel. But, it will complete after only finitely many such parallel moves.
I would expect it cannot in general require only finitely many Reidemeister moves. At least, the Reidemeister moves suggested above must be infinite in number since there are steps going from a diagram with infinitely many crossings to one with only finitely many. I doubt it's possible to use only finitely many Reidemeister moves except if $K_1$ is already known to be the unknot.
1) While the proof requires infinitely many Reidemeister moves to be performed, it's a bit different from a supertask because it's possible to do all of them in finite time by doing them in parallel. The diagram is split into countably many regions, at at each moment in time at most one Reidemeister move is performed in any given region.
2) Reidemeister's theorem does not apply since the object being isotoped is not a knot. However, all Reidemeister's theorem says is that if $D_1$ and $D_2$ are respectively diagrams for knots $K_1$ and $K_2$, then if $K_1$ is isotopic to $K_2$ there is a sequence of Reidemeister moves that transform $D_1$ into $D_2$. The swindle (when done diagrammatically) only needs the converse, that Reidemeister moves correspond to isotopies, which is true even in this setting. Furthermore, we don't need any theorems about existence of diagrams for wild knots since we provide the diagram ourselves.
3) Just to make what I said a bit clearer: Reidemeister's theorem is only used for converting isotopies into sequences of moves, but we can do the swindle by producing the program to perform infinitely many Reidemeister moves in finite time ourselves.
By the way, Mazur's swindle isn't necessary to prove that only the unknot has an inverse under connect sum. The reason is that Seifert genus satisfies
$$g(K_1\mathbin{\#}K_2) = g(K_1) + g(K_2)$$
and one can prove rather easily that $g(K)=0$ iff $K$ is the unknot.
There are some possible examples of "supertasks" in topology that rely on the way certain limits converge. One is in things like mapping telescopes for CW complexes (see Hatcher's book on algebraic topology). The weak topology lets you push problems away into higher dimensions, so to speak.
You might be able to consider Alexander's trick as a sort of supertask, but this is a stretch.
