Move CRTA offsets to kernel groups

[sidnadTT]

CRTA offsets are currently stored globally per dispatch_class, with a fixed maximum of 3 classes (DM0, DM1, COMPUTE). This doesn't scale very efficiently for Quasar, which has 8 DM cores + 4 TRISCs.
For Quasar's flexible programming model, the current design will waste L1 memory because it reserves CRTA space for all dispatch classes. We should move CRTA offset/size storage from ProgramConfig to KernelGroup. That way it's more dynamic: same kernel on multiple DMs shares just one CRTA L1 address and different kernels will get different CRTA offsets.