Generalized affine expression decomposition for SGPR promotion

[panditsa]

Add an N-way uniformity decomposition pass that splits any affine expression into additive components by symbol class (workgroup, induction variable, thread). This allows the LLVM backend to keep uniform contributions in SGPRs and fold them into hardware instruction fields (e.g. buffer_load soffset), reducing VGPR pressure.

Key changes:

• symbol_utils.py: add decompose_affine_by_uniformity(), a general N-way decomposition with cascade validation for cross-class terms
• emitter.py: add get_uniformity_classes() and gen_sympy_index_decomposed() so any handler can decompose arbitrary affine maps at the emitter level
• read_write.py: replace _split_index_three_way with the general utility, add _compute_linear_offset/_apply_uniform_offsets helpers, and apply the decomposition to handle_read, handle_write, and handle_gather_to_lds