[ENHANCEMENT]: Decouple Ollama from the Docker stack — opt-in bundled container + native-host mode

[vharkins1]

Description

The Docker dev stack always starts a bundled ollama container, and the app
service hard-depends on it (depends_on: ollama -> service_healthy). Consequences:

• The whole stack refuses to come up unless Ollama is healthy, even though only
  /forms/fill actually needs the LLM.
• There's no first-class way to run a host-native Ollama (Metal/GPU on
  macOS) — the bundled container can't use the Mac GPU.
• No supported "run without Ollama" path for working on non-LLM features
  (templates, weather/zipcode, Whisper transcription).

Proposed change

Make Ollama a swappable backend with two supported modes:

Bundled (default) — make up / make fireform

• Ollama runs in Docker, gated behind a bundled-ollama compose profile.

Native — make up-native

• Compose runs without the profile, so no Ollama container; the app reaches a
  host-native Ollama via host.docker.internal:11434.

Concretely:

• docker/dev/compose.yml: gate ollama behind profiles: ["bundled-ollama"];
  drop app's hard depends_on: ollama (app/services/llm.py already retries
  the connection); add extra_hosts: "host.docker.internal:host-gateway".
• Makefile: add up-native target + COMPOSE_NATIVE (omits the profile);
  default COMPOSE keeps the profile so existing behavior is unchanged.
• docker/.env.example: document bundled vs native OLLAMA_HOST.
• README.md: update run-from-source instructions.

Open question / follow-up

make up-native currently only prints a reminder to set
OLLAMA_HOST=http://host.docker.internal:11434 in .env.dev. If that edit is
missed, the app silently falls back to the container address and fills fail —
the same "fails silently depending on which path you follow" footgun as #382.
Consider having up-native export OLLAMA_HOST for its compose invocation so
native mode works without hand-editing.

Acceptance criteria

• make up starts the full stack incl. bundled Ollama (unchanged behavior)
• make up-native starts everything except Ollama and talks to a host-native instance
• Stack starts cleanly when Ollama is absent; only LLM-dependent endpoints degrade
• .env.example and README.md document both modes
• No regression to weather / zipcode / Whisper / template flows

Related: #382