Support prediction from pre-computed image embeddings

## Problem

Currently, the `predict()` methods on `TreeOfLifeClassifier`, `CustomLabelsClassifier`, and `CustomLabelsBinningClassifier` only accept raw images (file paths or PIL images). Internally, `predict()` always runs the full image encoding pipeline (`create_image_features()`) before computing classification probabilities.

This means that users who need **both** the image embeddings and the classification results must encode the image twice:

```python
classifier = TreeOfLifeClassifier(device='cuda')

# First pass: get embeddings for downstream use (e.g., similarity search)
features = classifier.create_image_features([classifier.ensure_rgb_image(img)])

# Second pass: predict re-encodes the same image internally
results = classifier.predict(images=img, rank=Rank.SPECIES)
```

The image encoding through the ViT backbone is by far the most expensive step in the pipeline. This doubles compute cost unnecessarily.

## Use cases

1. **Embedding + classification in one workflow**: Applications like [bioclip-image-search-lite](https://github.com/Imageomics/bioclip-image-search-lite) classify images and embed them for similarity search. With the current API, this requires encoding every image twice.
2. **Deferred classification**: A user pre-computes and stores embeddings (e.g., via the `bioclip embed` CLI), then later wants to classify them without needing the original images or GPU-heavy re-encoding.

## Proposed solution

Add an optional `image_features` parameter to the `predict()` methods. When provided, the method skips image encoding and computes classification probabilities directly from the supplied embeddings.

The method should validate the input embeddings before classification:
- Verify the tensor is 2D `(N, embedding_dim)`
- Check that `embedding_dim` matches the model's expected dimension (e.g., 768 for ViT-L/14)
- Normalize embeddings via L2 normalization before computing similarity, consistent with how BioCLIP 2 handles features in its forward pass

```python
# Encode once
features = classifier.create_image_features([classifier.ensure_rgb_image(img)])

# Predict from pre-computed embeddings — no re-encoding
results = classifier.predict(image_features=features, rank=Rank.SPECIES)
```

The change is fully backward-compatible — existing call signatures remain unchanged.

Start with the predict() methods on TreeOfLifeClassifier, CustomLabelsClassifier, and CustomLabelsBinningClassifier, then compare them with create_image_features() and the bioclip embed CLI flow. Done means pre-computed embeddings can be classified with shape and dimension validation, normalization, and no repeated image encoding while existing image-based calls remain supported.

Written by the indexing model from the issue text.