GridSearchOracle: persist _ordered_ids and _populate_next across reloads (#1055)

[SAY-5]

Closes #1055.

GridSearchOracle.__init__ builds two in-memory bookkeeping fields not present in Oracle:

• _ordered_ids: a LinkedList of trial ids in hp-combo order
• _populate_next: a queue of trial ids ready to produce their next combination

Neither is in Oracle.get_state / set_state and GridSearchOracle doesn't override them, so after a process restart with overwrite=False they reload empty. start_order rehydrates fine, the first end_trial pushes a trial id onto _populate_next, and the next populate_space calls _ordered_ids.next(old_trial_id) against an empty _data_to_index → KeyError: '<trial_id>' at gridsearch.py:80,197.

Override get_state / set_state on GridSearchOracle to persist both fields and rebuild the LinkedList from them on reload. Older state files (saved before this change) lazily rebuild from start_order so existing checkpoints keep working.

Regression tests in gridsearch_test.py cover:

• the state round-trip (write → fresh oracle → resume, no KeyError)
• legacy state files that lack the new keys still rehydrate via start_order

[gemini-code-assist]

Code Review

This pull request implements state persistence for GridSearch by adding get_state and set_state methods to the oracle, ensuring that search progress is correctly maintained across process restarts. It includes regression tests for state round-trips and backward compatibility with legacy checkpoints. Feedback was provided regarding the persistence of the LinkedList state, noting that directly saving the internal memory list may lead to corrupted trial ordering; a traversal-based approach was suggested to correctly capture the logical sequence of trial IDs.

[gemini-code-assist]

The _ordered_ids._memory list stores trial IDs in the order they were inserted into the LinkedList, which does not necessarily correspond to the logical sequence of the grid traversal. This happens when new hyperparameters are discovered or when using conditional search spaces, causing trials to be inserted in the middle of the list. Persisting _memory directly and then rebuilding the list by assuming its order is the sequence will result in a corrupted linked list upon search resumption. Instead, the linked list should be traversed using its next pointers to capture the correct logical sequence of trial IDs.

Suggested change

	state["ordered_ids"] = list(self._ordered_ids._memory)
	ordered_ids = []
	current_index = self._ordered_ids._next_index[None]
	while current_index is not None:
	ordered_ids.append(self._ordered_ids._memory[current_index])
	current_index = self._ordered_ids._next_index[current_index]
	state["ordered_ids"] = ordered_ids