Significant performance improvements and multi-GPU enhancements

New

3.4.56 Batch inference support : Improved concurrent inference requests from multiple deployments sharing a model using batching on DGPUs, reducing GPU busy time by 30% under multi-deployment load.
- For custom models, this is controlled by model advanced parameters (Custom models -> Advanced Configuration -> Parameters) : dynamic-batching-max-batch-size and dynamic-batching-batch-queue-window-us
3.4.56 Automatic Node fusion reduces number of discrete elements in a workflow to improve CPU utilization efficiency.
3.4.56 Experimental CUDA MPS support to further improve multi-gpu performance. This will become the default in upcoming releases. Enable using the following dev_config.toml
- ```
[lumeod]
enable_cuda_mps = true
```
3.4.56 The tensorrt-workspace-size-mib model parameter (Console "Advanced configuration / Parameters") sets the TensorRT builder workspace size in MiB used when converting an ONNX model to a Triton engine (default 64). Raise it for large backbones whose engine build fails with "could not find any implementation", which is more likely with dynamic batching enabled. A changed value reconverts the model on the next deployment start

Improved

3.4.56 RTSP output streams are now more reliable and reduce RTP packet loss
3.4.56 Multi-GPU gateways place a deployment on the GPU running the fewest deployments rather than the one with the most free memory, and each runner process is pinned to one GPU so that it opens a single CUDA context instead of one per GPU its deployments happen to use