[asm] Add pipelined double-buffering support with SGPR rotation

[harsh-nod]

Implement memref iter_arg handling for pipelined GEMM with g2s in the C++ WaveASM backend. When scf.for carries memref iter_args for double-buffering, the LDS base offsets are now materialized as SGPRs and rotated at the loop tail using s_mov_b32 swap sequences.

Key changes:

• RegionBuilder: detect LDS memref iter_args, resolve to SGPR offsets, propagate through block args, handle cross-swap at yield
• TranslateFromMLIR: use V_ADD_U32 directly with SGPR offsets in vector.load/store (V_MOV_B32 rejects SGPR sources)
• AMDGPUHandlers: handle dynamic SGPR-carried LDS base offsets in gather_to_lds m0 computation, prefer SALU when both operands are SGPRs
• LinearScanPass: fix block arg type propagation to use allocation mapping directly instead of condition iter_arg types (broken for cross-swap patterns)
• AssemblyEmitter: emit SGPR rotation copies at loop tail, detecting independent swap pairs and using 3-instruction swap with temporary