Leverage memory and network cost for dynamic task count decisions

Follow-up from #432.

In prepare_dynamic_plan.rs, calculate_cost produces a cost estimation with three dimensions — cpu, memory, and network — all of which are recorded as stage metrics (cpu_cost, memory_cost, network_cost). However, compute_based_task_count currently derives the task count from cost.cpu alone:

let compute_based_task_count = cost
    .cpu
    .get_value()
    .unwrap_or(&0)
    .div_ceil(nb_ctx.d_cfg.bytes_per_partition_per_second.max(1))
    .div_ceil(input_stage.plan.output_partitioning().partition_count())
    .clamp(1, nb_ctx.max_tasks()?);

We should explore how to also factor the memory and network cost estimates into the task count decision (e.g. a memory-pressure-driven scale-up, or accounting for network transfer cost), rather than relying on CPU cost only.

Start in prepare_dynamic_plan.rs by reading calculate_cost and compute_based_task_count, then trace how the cpu, memory, and network estimates become stage metrics. Determine and document how memory and network should influence task-count decisions, and consider the existing follow-up context in issue #432. Done means task-count selection no longer relies on CPU cost alone and the behavior is validated.