Existing tensor decoder has been bifurcated into two seperate gst
element as:
`yoloodv5tensordecoder`: decodes tensors output(masks) from detection-only
models e.g yolov8s.onnx
`yolsegv8tensordecoder`: decoder tensors output(masks and logits) from
segementation models e.g FastSAM or yolov8s-seg
YOLOv8 model have same tensor output format as FastSAM, so for better
generalization rename fastsamtensordecoder to yolotensordecoder. This
also requires code adaptation to support Yolo based model.
GstTensor contained two fields (data, dims) that were dynamicallay allocated. For
data it's for a GstBuffer and we have pool for efficient memory management. For
dims it's a small array to store the dimension of the tensor. The dims field
can be allocated inplace by moving it at the end of the structure. This will
allow a better memory management when GstTensor is stored in an analytics meta
which will take advantage of the _clear interface for re-use.
- New api to allocate and free GstTensor
To continue to support use-cases where GstTensor is not stored in an
analytics-meta we provide gst_tensor_alloc, gst_tensor_alloc_n and
gst_tensor_free that will facilitate memory management.
- Make GstTensor a boxed type
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/6000>
