RAG: Refactor Ollama lifecycle to lazy+persistent (remove auto-stop)

[sre-helmcode]

Description

BREAKING CHANGE: Refactor Ollama lifecycle management.

Current behavior: Ollama starts on team deploy (model_provider=ollama), stops when last such team stops (ref_count → 0).

New behavior: Ollama starts first time needed (RAG upload or Ollama team) and never stops automatically.

Changes:

• Remove ollamaMu mutex from server.go
• Remove ref_count increment in deployTeamAsync
• Remove ref_count decrement + StopOllama call in StopTeam
• Keep EnsureOllama, ConnectToNetwork, PullModel, WarmUp in deploy flow
• Keep DisconnectFromNetwork in StopTeam
• Update GetOllamaStatus handler (no more ref_count)

Why

RAG needs Ollama for embeddings at upload time AND query time. The old ref counting would stop Ollama when the last Ollama team stops, breaking RAG. Idle Ollama uses ~50MB RAM — negligible.

Acceptance Criteria

• Ollama no longer stops when last Ollama team stops
• Ollama still starts correctly for Ollama teams
• No regression in Ollama team deploy/stop flow
• All existing tests pass