fix(optimizers): fallback to seed_candidates in propose_new_candidates

[lelouar]

Summary

EvolutionaryOptimizer.propose_new_candidates calls select_candidate on best_candidates, which on the very first training step is still empty (it is populated later by maybe_add_candidate). The resulting None is then passed as selected_candidate into mutate_candidate / merge_candidate, crashing concrete subclasses such as OMEGA on selected_candidate.items() in omega.py.

The companion method on_batch_begin already handles this situation by falling back to a random seed candidate. This PR applies the exact same fallback in propose_new_candidates, keeping behaviour consistent across the class and letting OMEGA (and any future EvolutionaryOptimizer subclass) run fit() without pre-populating best_candidates.

Reproduction

program.compile(
    reward=synalinks.rewards.ExactMatch(in_mask=["answer"]),
    optimizer=synalinks.optimizers.OMEGA(
        language_model=lm,
        embedding_model=em,
    ),
)

await program.fit(x=x_train, y=y_train, epochs=1, batch_size=1, validation_split=0.2)

Before the fix this crashes at step 0 with:

AttributeError: 'NoneType' object has no attribute 'items'
  at omega.py:457 in mutate_candidate

Changes

• synalinks/src/optimizers/evolutionary_optimizer.py: in propose_new_candidates, when select_candidate(best_candidates) returns None, fall back to a random seed candidate (mirroring on_batch_begin).
• synalinks/src/optimizers/evolutionary_optimizer_test.py: add test_propose_new_candidates_falls_back_to_seed_when_best_empty. The test subclasses EvolutionaryOptimizer to record what mutate_candidate receives, and verifies it is the seed candidate (not None). The test fails on main and passes with this fix.

Test plan

• New regression test passes with the fix applied
• New regression test fails without the fix (verified by stashing the fix and re-running)
• Full evolutionary_optimizer_test.py suite passes (15/15)