refactor: move all llama code to models/llama out of meta reference (#1887)

# What does this PR do?

Move around bits. This makes the copies from llama-models _much_ easier
to maintain and ensures we don't entangle meta-reference specific
tidbits into llama-models code even by accident.

Also, kills the meta-reference-quantized-gpu distro and rolls
quantization deps into meta-reference-gpu.

## Test Plan

```
LLAMA_MODELS_DEBUG=1 \
  with-proxy llama stack run meta-reference-gpu \
  --env INFERENCE_MODEL=meta-llama/Llama-4-Scout-17B-16E-Instruct \
   --env INFERENCE_CHECKPOINT_DIR=<DIR> \
   --env MODEL_PARALLEL_SIZE=4 \
   --env QUANTIZATION_TYPE=fp8_mixed
```

Start a server with and without quantization. Point integration tests to
it using:

```
pytest -s -v  tests/integration/inference/test_text_inference.py \
   --stack-config http://localhost:8321 --text-model meta-llama/Llama-4-Scout-17B-16E-Instruct
```

llama_stack/cli/download.py

@@ -29,8 +29,8 @@ from rich.progress import (
 from termcolor import cprint
 
 from llama_stack.cli.subcommand import Subcommand
-from llama_stack.models.llama.datatypes import Model
 from llama_stack.models.llama.sku_list import LlamaDownloadInfo
+from llama_stack.models.llama.sku_types import Model
 
 
 class Download(Subcommand):

llama_stack/cli/model/describe.py

@@ -63,17 +63,6 @@ class ModelDescribe(Subcommand):
             ("Model params.json", json.dumps(model.arch_args, indent=4)),
         ]
 
-        if model.recommended_sampling_params is not None:
-            sampling_params = model.recommended_sampling_params.model_dump()
-            for k in ("max_tokens", "repetition_penalty"):
-                del sampling_params[k]
-            rows.append(
-                (
-                    "Recommended sampling params",
-                    json.dumps(sampling_params, indent=4),
-                )
-            )
-
         print_table(
             rows,
             headers,

llama_stack/cli/model/prompt_format.py

@@ -11,7 +11,7 @@ from pathlib import Path
 
 from llama_stack.cli.subcommand import Subcommand
 from llama_stack.cli.table import print_table
-from llama_stack.models.llama.datatypes import CoreModelId, ModelFamily, is_multimodal, model_family
+from llama_stack.models.llama.sku_types import CoreModelId, ModelFamily, is_multimodal, model_family
 
 ROOT_DIR = Path(__file__).parent.parent.parent
 

llama_stack/cli/model/safety_models.py

@@ -4,12 +4,12 @@
 # This source code is licensed under the terms described in the LICENSE file in
 # the root directory of this source tree.
 
-from typing import Any, Dict, Optional
+from typing import Any, Dict
 
 from pydantic import BaseModel, ConfigDict, Field
 
-from llama_stack.models.llama.datatypes import CheckpointQuantizationFormat, SamplingParams
 from llama_stack.models.llama.sku_list import LlamaDownloadInfo
+from llama_stack.models.llama.sku_types import CheckpointQuantizationFormat
 
 
 class PromptGuardModel(BaseModel):
@@ -23,7 +23,6 @@ class PromptGuardModel(BaseModel):
     is_instruct_model: bool = False
     quantization_format: CheckpointQuantizationFormat = CheckpointQuantizationFormat.bf16
     arch_args: Dict[str, Any] = Field(default_factory=dict)
-    recommended_sampling_params: Optional[SamplingParams] = None
 
     def descriptor(self) -> str:
         return self.model_id