[ML] Protect against multiple concurrent downloads of the same model

[davidkyle]

The tasks API is used to check for an existing model download when starting a model deployment but started tasks (or tasks about to start) are not immediately visible in the Tasks API leading to a race condition where successive calls to start model deployment may trigger multiple downloads. This is what happens in the Inference API where the default end points are used and multiple calls to inference will trigger the model download.

Download is a master node action so there is only one node in the cluster where the download can occur. Once we get to the download action and have access to the local taskManager checks for an existing download can be made more atomically. The change here is to look for a matching download task in the taskManager and if found register a listener on that tasks completion.