[Priyanshu-u07]

Fixes #436

#437 fixed the linker (deploys that wait for a new node). This covers the warm path: deploys onto a node that is already ready, where place_and_provision loads the model directly.

What changed in place_and_provision:

• RUNNING is written with update_state(..., unless_in=ENDED_STATES). If it's skipped because the deploy was terminated during the load, the model is unloaded on the worker and no endpoint is published.
• Both load-failure branches write FAILED with the same guard and release the GPU only if that write lands, in the same transaction. Before, a failed load after a terminate released the GPU a second time, freeing GPU another deploy on the node holds.
• The exception branch now records error_message, like the linker.

Testing:

• New test test_terminate_during_warm_load_is_not_overwritten (load ok and load failed): a terminate lands inside load_model on a node where another deploy holds 2 GPUs. It fails on main (RUNNING / FAILED over TERMINATED) and passes here; the other deploy's GPUs stay allocated.
• test_deploy_endpoint_pool_first.py is on CI's ignore list, so I ran it against a Postgres container with the repo schema. The other 22 failures in that file are the same on main (missing X-Organization-ID header since Check pool ownership and take new owners from the caller's org #417). Linker and worker lifecycle tests still pass.

Not changed: when RUNNING is skipped, the API response still reports DEPLOYING (the DB row is TERMINATED), same as the linker.