Adds reduce_sum and reduce_max operations

[zoecarver]

Adds reduce_sum and reduce_max operations following the same CB-input pattern as bcast and matmul.

• CB-input pattern: Reduce reads input tiles directly from CB, applies reduction, and writes to DST. Uses TTL_CBInputTileOpTrait for proper DST assignment handling
• Reduction dimensions: Supports row reduction (collapse rows), column reduction (collapse columns), and scalar reduction (collapse to single value). The dimension is encoded as an enum attribute and determines which TTKernel pool operation is emitted. Frontend API matches bcast dims.
• Multi-tile accumulation: When reducing across multiple input tiles, the lowering generates a loop that accumulates partial reductions. The scaler operand provides the reduction identity (0 for sum, -inf for max)
• Two-stage lowering: ConvertTTLToCompute builds the reduction loop structure with proper scaler handling. ConvertTTLTileOpsToTTKernel emits the appropriate reduce intrinsic based on reduction type and dimension

Self-contained and follows bcast's established lowering path and reuses the same CB lookup utilities. Risk is limited to reduce ops (adds new lowering, does not modify existing lowering patterns).