This is somewhat of a speculation, but I would assume that this is due to the modular way in which Sigma was designed. Namely, when Hugo Krawczyk designed Sigma, the main security property he was after was AKE security which basically consists of two things:
Session key indistinguishability: the adversary shouldn't be able to distinguish real session keys from random ones; and
Explicit entity authentication (EA): the property that once a protocol participant completes the protocol run, it is guaranteed that it was in fact communicating with the expected party, and that this party did indeed compute the same session key.
The EA property is basically a key confirmation step that assures liveness and authentication. This is achieved by calculating a MAC over some of the protocol data, using a key $K_m$ derived from the same master secret used to derive the session key.
The point is that you can achieve AKE security (that is, property 1 and 2) without any encryption at all! Indeed, when Krawczyk proves that Sigma satisfies AKE security (can't remember which paper this was; will try to find it later), he simply assumes that the encryption step is not there at all! (He also does this in his OPTLS paper, which is the precursor to TLSv1.3).
As I said, the standard security goal for most key exchange protocols have, since the original papers by Bellare and Rogaway, basically all been about AKE security. However, when Krawczyk designed Sigma he also wanted to add another, non-standard, feature, namely identity-protection. But given that he had already shown that the Sigma protocol without encryption achieved AKE security, it was a simple matter of adding encryption on top of this in order to also get identity-protection. Thus: MAC-then-Encrypt.
But notice that these two uses of MACing and encrypting are rather orthogonal and serve different purposes: the inner-MAC is supposed to provide EA security, while the outer encryption is supposed to provide identity-protection.
Also, note that Sigma typically does use Encrypt-then-MAC. In particular, in the instantiation of Sigma in IKEv2, the outer encryption is accompanied by an additional MAC outside of the encryption in standard EtM fashion. In IKEv2, the inner MAC key is called
SK_p*, the outer encryption key is called
SK_e* and the outer MAC key is called
* is either
r depending on whether this message is created by the initiator or the responder). Moreover, in newer instantiations of IKEv2 the outer encryption is replaced by dedicated authenticated encryption algorithms. In this case the
SK_a* keys are not used (and the whole EtM question becomes moot). In the TLSv1.3 instantiation of Sigma only authenticated encryption is used. But again, notice that inside the encryption (be it AE or EtM), there is an inner MAC whose purpose is to provide EA security.