You need Cosine theta, the power factor, to convert kVA to uniity power factor if your value is an inductive loading specification.
Overloading a generator could easily 'let the smoke out', especially if it is a modern, cheap generatorr with little overload capacity.
Ideally you wouldl need to load up the phases separately for test purposes, or is this machine a 'three wire' where three phases cannot be separated? I am suspecting this as your resistors are rated at 415 volts? Or maybe it is something else?
You need to wire them in parallel for a maximum power of 33kW, reducing in number to below your maximum load, whatever that might be (we don't know this value to within about 9kW - might be as low as 16 and as high as 25).
Your steps would be 3 x 3.7kW, 6 x 3.7kW or 9 x 3.7kW if you use you 3 switches to simply add in groups of 3 elements. The last option being patently inappropriate.
The third set could be switched in as a series set of three elements (at 3.7/3kW) or one plus two in parallel which would be approx 2.5kW (depending on actual resistance at the temperature achieved).
So, with your 3 switches you could achieve 11.1, 12.3, 22.2, 23.4, 24.7 or 33.3kW, but not all, as your last set could only be wired in one of the alternative configurations. All values approximate. Not a very useful set of values for testing purposes.
I see no other useful resistance values with your available components in the arrays that you are telling us you will be using. Unless of course you intend using 3 sets of 3 switches or 3 rotary switches with mutiple contacts.
The element sets in a series of three or the one and 'two in parallel would be expected to have much longer lifespans than the ones with 415 volts across them.
That lot said - now what exactly do you have, and want to do? Test each winding (phase)? Simply load up the system? 3 switches or 9?
RAB