TabFlex: Scaling Tabular Learning to Millions with Linear Attention