BF16 matmul regression on fourth-gen Xeon AMX tiles

The brief

BF16 matmul throughput regressed on fourth-gen Xeon AMX tiles after the blocking heuristic change. Retune the tile selection without hurting older generations.

Acceptance criteria

• Fourth-gen Xeon BF16 GEMM back within 2% of the previous release
• No regression on third-gen Xeon or client parts in benchdnn
• Heuristic change covered by a perf-guard test
• Tuning notes documented