Hotfix import of AzureMLOnBehalfOfCredential to only credential-less datastore scenario.

[alanpo1]

Hotfix for this PR here: #2929

We saw some issue with synapse cache using OLD cache dependency instead of NEW cache which included azure-ai-ml.

To remedy this need to move imports from new azure.ai.ml module to be inside of credential-less datastore code path to not impact old code path where we don't necessarily need to import from new conda module.

Note: for changing the try/catch from CredentialUnavailableError to an Exception here is an example job showing we still catch that error fine: https://ml.azure.com/runs/lime_branch_bz7qzy3b0s?wsid=/subscriptions/1aefdc5e-3a7c-4d71-a9f9-f5d3b03be19a/resourcegroups/momo-rampup-alanpo/workspaces/project-alantest-oai&tid=72f988bf-86f1-41af-91ab-2d7cd011db47#

[RichardLi1437]

have you tested with azure.ai uninstalled and the credential scenario works?

[alanpo1]

have you tested with azure.ai uninstalled and the credential scenario works?

I have a test running where I manually removed azure-ai-ml from the pip dependencies of all components and ran the create manifest e2e test. It is not entirely comprehensive of the changes made but it covers a large percent of the changes.

credential-less code-path Job w/o azure-ai-ml (expect failure): https://ml.azure.com/runs/good_bridge_yg74c7hr7d?wsid=/subscriptions/1aefdc5e-3a7c-4d71-a9f9-f5d3b03be19a/resourcegroups/momo-rampup-alanpo/workspaces/Project-alantest-oai&tid=72f988bf-86f1-41af-91ab-2d7cd011db47#

credential code-path job w/o azure-ai-ml (expect success): https://ml.azure.com/runs/upbeat_farm_vb0b49b3n2?wsid=/subscriptions/5fca341e-4ec3-45bd-b006-a7648d9376c5/resourcegroups/modelmonitoring-e2e-rg/workspaces/momo-e2e-test-ws-eastus&tid=72f988bf-86f1-41af-91ab-2d7cd011db47#