As stated in the second paragraph following the one you quote, the PDA being constructed is non-deterministic, which greatly simplifies the construction.
A non-deterministic machine is allowed to have points at which more than one action is possible, without any procedure for determining which one is correct. It is only necessary to prove that one possible action leads to correct termination.
Such machines are easy to construct in formal theory but building them in the real world is trickier. A back-tracking parser must explore alternatives in some order; if it chooses an alternative which fails to terminate, then it will never get around to trying another possibility.
In contrast, the non-deterministic machine terminates acts as if it took all possible branches at the same time. If any branch terminates in an accepting state, the machine can can terminate and accept the input. This can be simulated by making copies of the machine at every branch, and simulating the progress of all currently active copies by repeatedly advancing each one by one step. (More copies might be made at each step of each copy, so it can get a little closer unwieldy.)
Consequently, it does not matter to the PDA if there is left-recursion in the grammar. To be sure, the left-recursion means that some of the possible execution paths will never terminate, but as long as there is a path which terminates, the PDA will choose it.
There is a subtle but important point which I left out of the first version of this answer, which is that the definition of the language recognised by the PDA is the set of inputs for which the PDA terminates in an accepting state. Nothing guarantees that the PDA will eventually terminate on other inputs, and it is quite possible that the PDA will never halt if it is given a input which is not in the language it recognises. (This is another reason why non-deterministic PDAs aren't much use in practical parsers.)
This should not be surprising. If the machine were guaranteed to halt, then there would exist a machine which recognised the complement of the language. But if that were the case, the complement of every context-free language would be context-free, and we know that not to be true.