Export a pretrained embedding model (dinov2/dinov3) without a training run

Nobody has claimed this yet.

## Problem

To use a LightlyTrain backbone for embeddings — for example to feed LightlyStudio — you have to run a training job first. `export` and `embed` both require a checkpoint, and `pretrain` is the only way to produce one. There is no way to get an embedding model straight from a pretrained backbone name.

`load_model` returns a task model (detection/segmentation), not the embedding backbone. `package_helpers.get_wrapped_model(...)` + `train_helpers.get_embedding_model(...)` do build a pretrained `EmbeddingModel` with no training, but both are private (`_models`, `_commands`).

## Ask

A public API to export or load a pretrained embedding model from a model name, no training:

```python
lightly_train.export(out="model.pt", model="dinov2/vits14", part="embedding_model", format="torch_model")
# or
model = lightly_train.load_embedding_model("dinov2/vits14") # torch.nn.Module
```

Cover the families users reach for first: `dinov2/*` and `dinov3/*`.

## Why

Users who just want good embeddings now should not have to run a 1-epoch training job as a stand-in. The LightlyStudio embeddings tutorial (lightly-ai/lightly-studio#2160) currently does exactly that. A one-liner to drop a stock backbone into an embedding pipeline makes the LightlyTrain → LightlyStudio path feel like a few lines of code.

Start by tracing the public export, embed, pretrain, and load_model entry points, then inspect package_helpers.get_wrapped_model(...) and train_helpers.get_embedding_model(...). Define how a model name selects dinov2/* and dinov3/* backbones without training, expose the chosen API, and verify that it returns or exports an embedding model in the requested format.

Written by the indexing model from the issue text.