titvm: Add TVM inference plugin (titvm) for AM62D

[PrathamTI]

Implement TVM-based inference element with TI C7x DSP acceleration using the TVM runtime with TIDL backend for TI AM62D platform.

Current Implementation:

Binary tensor inference support

• Accepts raw float32 tensor input (application/octet-stream)
• Produces float32 tensor output (application/octet-stream)
• Direct TVM C++ runtime API integration

Performance benchmarking

• Configurable inference iterations (iterations property)
• Min/avg/max/fps statistics (benchmark property)
• First inference includes model loading overhead
• Subsequent inferences show true runtime performance

Model loading

• Loads TVM artifacts: deploy_lib.so, deploy_graph.json, deploy_param.params
• Artifacts must be compiled for AM62D C7x DSP with TVM TIDL backend
• model-path property specifies artifacts directory

Properties:

• model-path: Path to TVM artifacts directory (required)
• iterations: Number of inference runs (default: 1)
• benchmark: Enable performance statistics (default: true)

Pipeline Example:
gst-launch-1.0 filesrc location=input_tensor.bin blocksize=1048576 ! \ titvm model-path=/usr/share/tvm/artifacts iterations=1 ! \ filesink location=output_tensor.bin

TODO:

Audio enhancement pipeline integration

• Add audio-mode property for complex spectral data processing
• Implement deinterleave/interleave operations via C7x RPMsg
• Integrate with STFT/ISTFT preprocessing/postprocessing elements
• Add fft-size and input-shape properties for audio models