Migrate AICore kernel compilation from C++ to Python and simplify DeviceRunner API

[ChaoWao]

This commit fundamentally restructures how AICore kernels are compiled and loaded, moving kernel compilation responsibility from C++ to Python and simplifying the device runtime API.

Key changes:

Kernel Compilation Migration:

• Extract kernel compilation from C++ (kernel_compiler.cpp/h) to Python
• Add PTOCompiler class for Python-side kernel compilation using ccec
• Add elf_parser.py for extracting .text sections from ELF object files
• Remove binary_loader.cpp/h and kernel_compiler.cpp/h (replaced by Python)
• Kernels now compiled and registered at Python layer before runtime execution

DeviceRunner API Simplification:

• Replace DeviceRunner Init/Run/Finalize with set_device() and launch_graph()
• Remove pto_isa_root parameter (no longer needed after moving compilation to Python)
• Add RegisterKernel() API for kernel binary registration
• DeviceRunner now auto-finalizes via destructor (no manual Finalize needed)
• Device initialization split into set_device() (early) and lazy binary loading (in launch_graph)

Runtime API Changes:

• Change Runtime allocation from C++ new/delete to placement new pattern
• Add GetGraphSize() for user-side buffer allocation
• Rename ValidateGraph() to FinalizeGraph() to better reflect responsibility
• Remove function_cache.h (kernel lookup now uses direct registration)

AICore Executor Abstraction:

• Extract aicore_executor.h with core type and device args abstraction
• Add DeviceArgs struct to replace uint64_t kernel parameters
• Refactor kernel.cpp to use new executor pattern

Documentation and Testing:

• Update example/main.py to demonstrate new Python-first workflow
• Simplify README testing instructions (use ./ci.sh)
• Add ci.sh for streamlined testing

Net result: -270 lines, cleaner separation of concerns, Python has full control over kernel compilation and registration lifecycle.