Create Fanout

POST /jobs/fanouts: 2 to 8 jobs trained together on one dataset version so their models can be compared, e.g. one per base. allocation_mode: separate (the default) gives each job its own GPUs; shared_lora trains LoRA SFT jobs of one base model as sessions on one set of GPUs holding that base once, each job with its own adapter, data, logs, receipt and budget (lora.budget_h100_hours, required), billed its share of the GPU time: the shares add up to the allocation's GPU time exactly.