Here is a little context to maybe complement Tom's and Nick's answers.
The definition (2) in terms of being flat over $\pi_* \Bbb S$ is new - it's a specialization of a definition of flatness over $R$ that comes from derived / spectral algebraic geometry. Over $\Bbb S$ this is so rare that most examples are either in characteristic zero, or they are tailor-made to satisfy it (eg: localizations). The main utility would be to get a Kunneth theorem, identifying $E_* X$ with $E_* \otimes_{\pi_* \Bbb S} \pi_* X$. It has very specific goals, such as making base-change and descent effectively calculable.
As Nick says, Ravenel's interest is in the general Adams-Novikov spectral sequence, and specifically getting better than the $E_1$-term. (In any case where the version (2) definition applies, the Adams-Novikov spectral sequence is pretty degenerate.) Slightly more explicitly, the $E_1$-term starts with the homotopy groups of $E \wedge X$, $E \wedge E \wedge X$, and so on, and produces a spectral sequence for the homotopy groups of $X$. We would ideally like to understand $\pi_*(E \wedge E \wedge \dots \wedge E \wedge X)$ in terms of $E_* X$.
This leads to definitions (1) and (3). Both of these are geared at getting a Kunneth isomorphism $$\pi_*(E \wedge E \wedge Y) \cong \pi_*(E \wedge E) \otimes_{\pi_* E} E_* Y$$ which we can then apply inductively to $X$. Definition (1) gets at this Kunneth isomorphism explicitly, because the splitting of $E \wedge E$ tells you that $$ E \wedge E \wedge Y \simeq \bigvee_i \Sigma^{n_i} (E \wedge Y). $$ Definition (3), by contrast, gets at this isomorphism with the Kunneth spectral sequence. You re-express $$ E \wedge E \wedge Y \simeq (E \wedge E) \wedge_E (E \wedge Y) $$ and this has a spectral sequence $$ Tor_{**}^{E_*} (E_* E, E_* Y) \Rightarrow \pi_*((E \wedge E) \wedge_E (E \wedge Y)). $$ Definition (3) implies that this degenerates to a Kunneth isomorphism.
So why would we choose version (1), when version (3) is usually strictly stronger? There are two reasons.
Maybe $E$ isn't good enough. Version (3) has an assumption - it only applies if $E$ is a highly structured (associative) ring spectrum, so that modules over $E$ and smash products over $E$ make sense. This is a problem if you're working with a ring spectrum that doesn't admit that nice of a multiplication. By contrast, version (1) applies to not-asssociative-but-homotopy-associative things -- for example, the mod-$p$ Moore spectrum when $p$ is a prime greater than 3.
More seriously, maybe it is 1992, when the orange book is being published. In 1992, categories of module spectra over highly structured ring spectra don't really exist, and there is no high-powered Kunneth spectral sequence available to you. Version (1) is what you've got.