feat: Remove global (cross-worker) stopping criteria

[nastaran78]

This PR restructures how stopping conditions are handled in NEPS by removing the old global, cross-worker logic and shifting all controlling parameters to the level of individual workers.

definition: A worker_id represents a single execution of neps.run. Running NEPS again—sequentially or in parallel—creates a new worker_id. Stopping parameters are scoped to that worker rather than shared across all active workers.

Previously, fidelity usage and total evaluations were tracked also in memory of the worker, and trial files on disk were not checked for a single worker. For avoiding inconsistency and having one single source of truth, this PR moves to a disk-based ground truth and removes the now-invalid parameters from the worker settings, e.g. max_evaluation_for_worker, max_cost_for_worker,..

The earlier design also treated fidelity_to_spend as a global limit across workers, which made it impossible to define per-worker limits, especially in multi-fidelity setups where evaluations_to_spend could not be combined with that system. cost_to_spend and evaluations_to_spend were effectively the same logic pretending to be different knobs.

Now each worker has its own independent stopping parameters, and multiple stopping criteria are supported instead of being restricted to just one.

One trade-off: when users pass something like evaluations_to_spend, they must be aware of how much has already been consumed by the worker and adjust the value accordingly after failures or restarts. This will be addressed by adding a new summary report—total evaluations, total evaluation time, total cost, total fidelity usage.

This brings the API closer to predictable, per-worker resource accounting while cleaning up several historical inconsistencies.