Simplify backfill window to always return full projection

[JaeseungYeom]

• Remove the full_horizon option across the C++, Python, and gRPC APIs.
• Always include projected releases for all running jobs and update the dispatcher, tests, and documentation accordingly.

Changes in the current worktree:

• Scheduler API: adds on-demand waiting_resource_area() and shared prediction_horizon() logic.
• Existing FCFS implementations: scan their existing queue entries only when prediction is requested.
• Backfill window: always returns all projected running-job releases; removes full_horizon.
• gRPC/Python: exposes prediction-horizon queries.
• Dispatcher: adds performance-aware multi-system routing.
• Tests/CMake: adds dispatcher, standard-FCFS, Python, and gRPC coverage.
• Documentation: describes the new APIs and dispatch workflow.

Runtime qualifications:

• Normal scheduling has no new ongoing bookkeeping.
• Prediction queries cost roughly O(Q + R log R) with temporary O(R) storage.
• get_backfill_window() also constructs its release map only when queried.
• FCFS-alt now writes its reservation into the already-existing m_fcfs_reservation_time field instead of a local variable. That is a correctness fix, not new state.