feat(backend): retire deprecated LLM models with family-aware migration (#13089)

### Why / What / How

**Why:** The `LlmModel` enum in `backend/blocks/llm.py` still contained
entries that providers have retired or superseded. Removing them was
unsafe because the boot-time safety net (`migrate_llm_models` in
`backend/data/graph.py`) force-mapped any out-of-enum value to
`DEFAULT_LLM_MODEL` (currently `GPT5_2`). That meant a Claude Opus user
would silently end up on GPT-5.2 — different provider, different price
tier, different behavior.

**What:** Retire 24 deprecated `LlmModel` members in one PR with:
1. a Prisma migration that rewrites stored model values (in both
`AgentNode.constantInput` and `AgentNodeExecutionInputOutput.data`) to
same-family replacements;
2. a `LEGACY_MODEL_MAPPINGS` dict that the safety net consults before
falling back to the global default;
3. removal of the dead entries from `MODEL_METADATA` / `MODEL_COST` /
`TOKEN_COST` and any pinned model strings in tests.

**How:**
- `LEGACY_MODEL_MAPPINGS: dict[str, LlmModel]` lives in
`backend/blocks/llm.py` as the single source of truth. The migration SQL
and the safety net both rewrite from the same mapping.
- `migrate_llm_models` now runs two SQL passes per LLM field: one
targeted `UPDATE` per legacy mapping (family-aware), then the existing
catch-all that maps anything still out-of-enum to `DEFAULT_LLM_MODEL`.
- AI/ML API stragglers (Qwen 2.5-72B, nvidia/llama-3.1-nemotron,
Meta-Llama-3.1-70B Turbo, Llama-3.2-3B Turbo) have no direct same-family
successor on AI/ML's current catalogue, so they all map to
`meta-llama/Llama-3.3-70B-Instruct-Turbo` (kept in the enum). This is
called out in the migration header.

Closes
[OPEN-3109](https://linear.app/autogpt/issue/OPEN-3109/retire-deprecated-llm-models-with-family-aware-migration).

### Changes 🏗️

- **New migration**
`migrations/20260512120000_retire_deprecated_llm_models/migration.sql` —
24 family-aware rewrites against `AgentNode.constantInput` and
`AgentNodeExecutionInputOutput.data` (preset overrides).
- **`backend/blocks/llm.py`** — removed 24 deprecated `LlmModel` members
+ their `MODEL_METADATA`; added `LEGACY_MODEL_MAPPINGS` shared with the
migration and the safety net.
- **`backend/data/graph.py`** — `migrate_llm_models` now does a targeted
`UPDATE` per legacy mapping before the catch-all fallback.
- **`backend/data/block_cost_config.py`** — dropped stale `MODEL_COST` +
`TOKEN_COST` entries for the removed models.
- **Tests** — updated pinned deprecated model strings in
`blocks/test/test_llm.py`, `data/credit_test.py`, `util/prompt_test.py`,
`copilot/config_test.py`, `copilot/model_normalize_test.py`, and 5
`copilot/sdk/*_test.py` fixtures.

#### Retired → replacement (verified against current provider docs)

| Retired | Replacement |
|--|--|
| `claude-3-haiku-20240307` | `claude-haiku-4-5-20251001` |
| `claude-opus-4-20250514` | `claude-opus-4-7` |
| `claude-sonnet-4-20250514` | `claude-sonnet-4-6` |
| `claude-opus-4-1-20250805` | `claude-opus-4-7` |
| `gpt-4-turbo` | `gpt-4.1-2025-04-14` |
| `o1` | `o3-2025-04-16` |
| `o1-mini` | `o3-mini` |
| `google/gemini-2.5-pro-preview-03-25` | `google/gemini-2.5-pro` |
| `google/gemini-2.5-flash-lite-preview-06-17` |
`google/gemini-2.5-flash` |
| `cohere/command-r-08-2024` | `cohere/command-a-03-2025` |
| `cohere/command-r-plus-08-2024` | `cohere/command-a-03-2025` |
| `mistralai/mistral-nemo` | `mistralai/mistral-small-3.2-24b-instruct`
|
| `microsoft/wizardlm-2-8x22b` | `microsoft/phi-4` |
| `moonshotai/kimi-k2` | `moonshotai/kimi-k2.6` |
| `moonshotai/kimi-k2-0905` | `moonshotai/kimi-k2.6` |
| `z-ai/glm-4-32b` | `z-ai/glm-4.6` |
| `z-ai/glm-4.5` | `z-ai/glm-4.6` |
| `z-ai/glm-4.5-air` | `z-ai/glm-4.7-flash` |
| `z-ai/glm-4.5-air:free` | `z-ai/glm-4.7-flash` |
| `z-ai/glm-4.5v` | `z-ai/glm-4.6v` |
| `Qwen/Qwen2.5-72B-Instruct-Turbo` |
`meta-llama/Llama-3.3-70B-Instruct-Turbo` |
| `nvidia/llama-3.1-nemotron-70b-instruct` |
`meta-llama/Llama-3.3-70B-Instruct-Turbo` |
| `meta-llama/Meta-Llama-3.1-70B-Instruct-Turbo` |
`meta-llama/Llama-3.3-70B-Instruct-Turbo` |
| `meta-llama/Llama-3.2-3B-Instruct-Turbo` |
`meta-llama/Llama-3.3-70B-Instruct-Turbo` |

### Out of scope

- `DEFAULT_LLM_MODEL` is unchanged.
- Stagehand's `StagehandRecommendedLlmModel` subset references only
Claude 4.5/4.6 Sonnet (both kept) — untouched.
- No new models added. This PR is purely retirement + safety-net
hardening.

### Checklist 📋

#### For code changes:
- [x] I have clearly listed my changes in the PR description
- [x] I have made a test plan
- [x] I have tested my changes according to the test plan:
- [x] `poetry run pytest
backend/data/graph_test.py::test_migrate_llm_models_uses_schema_prefix_placeholder`
— passes (regression: migrate_llm_models still uses the
`{schema_prefix}` placeholder for all queries, including the new
targeted ones).
- [x] `poetry run pytest backend/blocks/test/test_llm.py` — 60 tests,
all green.
- [x] Local `docker compose up -d` boot check — backend starts cleanly
and the boot-time safety net runs without errors (pending — please
verify in CI / before merge).
- [x] Apply migration against a staging snapshot of a database that
contains a graph pinned to one of the retired models, confirm
`constantInput.model` is rewritten to the mapped replacement (not
`DEFAULT_LLM_MODEL`).

#### For configuration changes:

- [x] `.env.default` is updated or already compatible with my changes —
no env changes.
- [x] `docker-compose.yml` is updated or already compatible with my
changes — no compose changes.
- [x] I have included a list of my configuration changes in the PR
description (under **Changes**) — n/a.