An intuitive account of this equivalence can be given over a sample domain (a set of numbers, for instance), while we ought to rely first and foremost on derivations. Here is one in axiomatic (Hilbert) style:
$1.\qquad\forall x\forall yP(x, y)\qquad\qquad\qquad\qquad\text{hypothesis}$
$2.\qquad\forall x\forall yP(x, y)\rightarrow\forall yP(x, y)\qquad\,\text{Axiom 4}$
$3.\qquad\forall yP(x, y)\qquad\qquad\qquad\qquad\quad\text{Modus ponens 1, 2}$
$4.\qquad\forall xP(x, y)\rightarrow P(x, y)\qquad\qquad\;\text{Axiom 4}$
$5.\qquad P(x, y)\qquad\qquad\qquad\qquad\qquad\text{Modus ponens 3, 4}$
$6.\qquad\forall xP(x, y)\qquad\qquad\qquad\qquad\quad\text{Generalisation from 5}$
$7.\qquad\forall y\forall xP(x, y)\qquad\qquad\qquad\qquad\text{Generalisation from 6}$
By symmetry, $\forall x\forall yP(x, y)$ is derived from $\forall y\forall xP(x, y)$ as well and we conclude
$$\forall x\forall yP(x, y)\leftrightarrow\forall y\forall xP(x, y)$$
Remark. Each axiomatic system with its axioms and inference rules just as stated should be applied in its entirety; nothing more and nothing less. In the system I've used here, each line (including hypothesis) is accepted as tautology. Thus, line 5 states that $P(x, y)$ is true for arbitrary $x$ and $y$. Hence, generalisation is not an illegitimate binding of a variable.
Axiom 4 of this system is $\forall x\phi(x)\rightarrow\phi(t)$ where $t$ is free for $x$ in $\phi$. Clearly, each variable is free for itself.