Support targeting specific instances

[peterschmidt85]

Summary

Adds a strict instances schema for runs so users can target existing fleet instances explicitly:

• short form: instances: [my-fleet-1]
• by name: instances: [{name: my-fleet-1}]
• by hostname: instances: [{hostname: worker-1}]
• by fleet instance number: instances: [{fleet: my-fleet, instance: 1}], including project/fleet

Compared to #3925, this PR keeps the normal fleet selection path mostly unchanged. The new selector resolution and targeted reusable-instance assignment path is used only when instances is set; missing targets fail without provisioning new capacity.

Testing

Automated:

• uv run --no-sync pytest src/tests/_internal/core/models/test_profiles.py src/tests/_internal/server/services/test_instances.py src/tests/_internal/server/services/runs/test_plan.py src/tests/_internal/server/background/pipeline_tasks/test_submitted_jobs.py -q (100 passed, 77 skipped)
• uv run --no-sync ruff check ...
• uv run --no-sync pyright -p .
• git diff --check

E2E:

• Backend AWS fleet with two t2.small instances:
  • normal run without instances
  • short-form targeted instance
  • missing target leaves the fleet at two idle instances and does not provision
• SSH fleet backed by two EC2 instances in one VPC/security group using the dstack AMI:
  • normal run
  • target by name
  • target by hostname using a service
  • target by {fleet, instance}
  • missing target fails without occupying instances
  • multinode negative: one selected physical instance is not enough
  • multinode positive: two selected physical instances both run one node

All E2E AWS resources were cleaned up after verification.

AI Assistance: implemented and tested with Codex.