feat: top-K winner-takes-most merge prediction rewards

[bittoby]

Summary

• Replace proportional reward distribution across all positive-EMA miners with a top-3 winner-takes-most system (50%/35%/15% split of the 15% prediction emissions)
• Add PREDICTIONS_TOP_K and PREDICTIONS_TOP_K_SHARES constants to configure the distribution
• Ties broken by rounds (more settled issues = higher rank); unallocated shares when fewer than 3 eligible miners are not redistributed

Changes

• gittensor/constants.py — added PREDICTIONS_TOP_K = 3 and PREDICTIONS_TOP_K_SHARES = [0.50, 0.35, 0.15]
• gittensor/validator/forward.py — rewrote build_prediction_ema_rewards() to rank miners by (ema_score, rounds) descending and award only top-K with fixed shares
• tests/validator/merge_predictions/test_merge_predictions.py — added 8 tests covering standard split, edge cases (fewer miners, ties, deregistered miners, empty EMA), and total reward cap
• tests/validator/test_emission_shares.py — added guard-rail tests for top-K shares summing to 1.0 and length matching PREDICTIONS_TOP_K

Test plan

• pytest tests/validator/test_emission_shares.py -v — constant guard-rails pass
• pytest tests/validator/merge_predictions/test_merge_predictions.py -v — all 8 top-K reward distribution tests pass
• Verify on testnet that only top-3 miners by EMA receive prediction rewards with correct share allocation

Closes #289

[bittoby]

@anderdc Could you pls review this PR?