Refactor: unify register access into generic read_reg/write_reg interface

[ChaoWao]

Replace purpose-specific register functions (read_task_id_by_reg, set_aicore_busy/idle, enable/disable_aicore_register, write_task_id_to_aicore, read_aicore_status) with a generic read_reg/write_reg API parameterized by RegId enum.

Platform layer:

• Add RegId enum (DATA_MAIN_BASE, COND, FAST_PATH_ENABLE) and reg_offset() mapping in platform_config.h
• Add SIM_REG_BLOCK_SIZE constant for simulation register memory sizing
• Rewrite a2a3 aicpu_regs.cpp to use read_reg/write_reg with MMIO pointers
• Add a2a3sim aicpu_regs.cpp with identical interface over host memory
• Rewrite a2a3 inner_kernel.h: replace inline asm helpers with read_reg/write_reg dispatching on RegId via switch
• Add a2a3sim inner_kernel.h register simulation: thread-local g_sim_reg_base, read_reg/write_reg over simulated memory, pipe_barrier macro
• Update aicpu_regs.h header to declare unified read_reg/write_reg interface

Simulation runner:

• Allocate per-core simulated register blocks in DeviceRunner::run()
• Pass physical_core_id through aicore_execute_wrapper to set up thread-local g_sim_reg_base before kernel execution
• Update function pointer signatures to include physical_core_id

Runtime layer:

• Add physical_core_id field to Handshake struct (aicpu_build_graph, tensormap_and_ringbuffer)
• Update aicore_executor.cpp to use write_reg/read_reg instead of set_aicore_idle/busy/read_task_id_by_reg
• Update aicpu_executor.cpp: replace all register helper calls with read_reg/write_reg, add pre-set COND=BUSY before task dispatch to prevent false completion detection from stale IDLE state