[metal] MslAstWalker + MetalBackend MSL codegen + device function hook + tests

[aditvenk]

Stacked PRs:

• [metal] Support elementwise kernels with >1D tensors #1765
• ->[metal] MslAstWalker + MetalBackend MSL codegen + device function hook + tests #1764
• [metal] Metal codegen for load/store/mask_to #1763

---

[metal] MslAstWalker + MetalBackend MSL codegen + device function hook + tests

• Add helion/_compiler/metal/ package:
  • msl_ast_walker.py: MslAstWalker class with _generate_elementwise()
    and AST-to-MSL C++ converter functions. Handles tl.cast, tl.where,
    tl.full, tl.reshape, tl.sigmoid, tl.sqrt_rn, libdevice.*,
    triton_helpers.maximum/minimum, select, ast.If for masked stores.
• Modify codegen_function_def to post-process for Metal: standard path
  runs first (variable renames, constexpr inlining, scalar preambles),
  then MetalBackend.generate_msl_function extracts the FunctionDef body,
  creates an MslAstWalker, and replaces with a zero-arg function
  returning (msl_source, kernel_name).
• Add MetalBackend methods: transform_host_arg (wraps scalar args as
  float32 1-element buffer tensors), build_launcher_args, inline_constexpr,
  generate_msl_function, and expression generators (program_id_expr,
  cdiv_expr, sympy_printer_expr, grid_index_expr, inductor_op_overrides)
• In a subsequent change, we will avoid passing scalar args as 1-element
  buffer tensors, and instead bake them directly into the MSL source.
• 26 tests: arithmetic (add/sub/mul/div/neg), scalar args (saxpy),
  activations (relu/silu/gelu), math ops (exp/log/sqrt/abs/sin/cos/clamp),
  dtypes (float16/bfloat16/int32), bounds masking (OOB sentinel checks,
  codegen mask assertions)

[aditvenk]

Will shortly open a new PR that refactors the commits in the stack differently