@Simboubou: Nope, that's a design decision in CRe (the ePub engine used by KOReader). The trade-off is that page rendering is then much easier on the CPU (i.e., in the perfect scenario: performance hit once ([first] book open), performance gains many (navigating the book)).
It's a trade-off I'm generally happy with (although we are interested in degenerate cases where stuff is taking an inordinately large amount of time, c.f., the KOReader forum), although it's largely historical: CRe was designed way back when, when embedded devices had ridiculously low specs (... they still do. But it used to be way worse
).
Feel free to try Plato, which probably does chunked parsing/decoding.