MLIR `linalg` Dialect Lowering for Matrix Multiplication

Right now, benchmark hardware dispatch testing (e.g. tests/backend/pass/plugin_npe.vx) uses a dummy flat C-ABI dispatch test via spawn on(Topology::ANE).

Objective:
To officially wire the standard tensor::matmul operations inside spawn on blocks to lower directly to proper linalg.matmul MLIR operations, allowing proper pass handoff to hardware vendor plugins.

Tasks:

• Update the MLIR code generator to recognize matrix multiplication inside spawn on regions.
• Ensure the emitted IR uses linalg.matmul with the proper memref indices and memory spaces.
• Validate that the NPU plugin can successfully ingest and lower this linalg block to native NPU commands.

Start with tests/backend/pass/plugin_npe.vx and trace how tensor::matmul inside spawn on(Topology::ANE) reaches the MLIR code generator. Compare the emitted IR with the expected linalg.matmul form, including memref indices and memory spaces, then validate that the NPU plugin ingests and lowers the resulting block to native NPU commands.