Feature/forecast

[Oklyagina]

• train injection forecasters against the true t+12 target
• construct one shared forecasted future grid state per observation
• simulate each selected line contingency independently
• build forecaster features from the post-contingency operating state
• generate labels by applying the configured agent to the contingency state
• skip invalid future power-flow states instead of treating them as agent failures
• keep training agent-agnostic through the existing act/factory interface
• remove the RL agent dependency from runtime HGB inference
• add multi-contingency prediction with probability and binary output per line
• update the example, documentation, and synthetic tests