Hello LuĆs,
in an MLAG setup the switch of the MLAG pair receiving a frame for a device connected via MLAG will always send it directly, unless the link has failed. This is different from the default settings of S-Series VSB, for example.
LACP should be used for all LAGs if possible (e.g. correct VMware license level).
As paul wrote, link selection is a hop-by-hop process. I use L3_L4 hashing unless I have a really good reason to use something else (e.g. L3_L4 not supported or non-IP traffic).
The hop-by-hop load sharing combined with using the local MLAG port for delivery combined result in very low ISC usage by default, unless single attached devices connect to an MLAG pair. Frames arriving at the wrong switch need to traverse the ISC.
Best regards,
Erik