Perhaps a better-known name for this technique is arithmetization.
The broad idea behind it is to encode boolean logic into low-degree polynomials, which (for certain technical reasons) have various useful analytical techniques available (probably most obviously the Schwartz-Zippel lemma).
This is in particular useful for proving soundness of interactive proofs, and was key to Shamir's proof of $\mathsf{IP} = \mathsf{PSPACE}$.
$f$ is a map - however I think it's also a polynomial. It's probably in the form of a polynomial of maximum degree $v$ & it takes it's coefficients from $\mathbb{F}_2$.
Even if you can think of $f$ as a polynomial, it is better not to.
We have some initial "standard" computational object we want to analyze, which typically means
- a boolean formula, or
- a boolean circuit, or
- a truth table.
You can view these all as polynomials over $\mathbb{F}_2$ (and arguably that is precisely what boolean circuits are), but sometimes you'll have a formula instead or whatever.
The idea behind finding an extension polynomial (or, as I said before, appealing to "arithmetization") is to encode this (standard) object as a polynomial $g(x) \in \mathbb{F}[x_1,\dots,x_v]$ that "agrees with" $f(x)$ in the sense that
$$\forall (x_1,\dots,x_v)\in\{0,1\}^v : f(x_1,\dots,x_v) = g(x_1,\dots,x_v).$$
This can easily be done for certain operations, for example $g_{AND}(x,y) = xy$ is the extension of AND, $g_{NOT}(x) = 1-x$ is the extension of NOT.
For other operations it is a little less straightforward, for example
$$g_{XOR}(x,y) = x+y-2xy = 1 - (xy - (1-x)(1-y)).$$
is the arithmetization of XOR (I think), and is perhaps not obvious beforehand.
In the comments you've asked why we care.
Perhaps the best motivation is Schwartz-Zippel Lemma, but its s technical lemma who's utility may not be useful immediately.
The "elevator pitch" for arithmetization is
- it was key to proving $\mathsf{IP} = \mathsf{PSPACE}$ (via Shamir's "sumcheck" protocol), one of the first big results in interactive proof systems, and
- the proof of this result was very special (non-relativizing, and not a natural proof). Arguably "arithmetization" is one of ~3 or 4 fundamental proof techniques we know in complexity theory. Until 2009, it was really the main "big" techniques which still had a shot at showing that $\mathsf{P} \neq \mathsf{NP}$ --- it no longer has a shot though.
Anyway, for an explicit book reference, I know it is at least contained in Arora & Barak's Computational Complexity : A Modern Approach. Using ctrl+f to search "arithmetization" immediately brings you to chapter 8.5.2, which discusses the approach.
In general, searching on this term will likely be much more fruitful than "extension polynomial", which may have more inadvertent name collisions.