X0 pred

[adsfssd]

https://arxiv.org/pdf/2511.13720

adds:

  --jit_loss_type {v,x0,eps}
                        Loss space to use according to JIT
  --jit_pred_type {v,x0,eps}
                        Model prediction space to use according to JIT
  --jit_loss_weights x0_mult v_mult eps_mult
                        Perform weighted average across loss types
  --jit_t_eps JIT_T_EPS
                        t_eps to avoid division errors when converting loss types, ignored if implicit scale is used
  --jit_implicit_scale  Calculate loss in a scaled space where possible to avoid division instability

to sdxl_train.py