docs: update semdedup docs for vLLM default backend (PR #1606)

[lbliii]

Description

Updates v26.04 fern documentation to reflect the vLLM default embedding backend change from #1606. Replaces all EmbeddingCreatorStage references with VLLMEmbeddingModelStage, updates the default model from sentence-transformers/all-MiniLM-L6-v2 to google/embeddinggemma-300m, removes stale parameters (embedding_max_seq_length, embedding_padding_side, embedding_pooling, embedding_model_inference_batch_size), and adds new vLLM-specific parameters (embedding_pretokenize, embedding_vllm_init_kwargs, model_cache_dir). Replaces the 26.02 release notes with a 26.04 skeleton including the vLLM default entry and breaking changes.

Usage

from nemo_curator.stages.text.deduplication.semantic import TextSemanticDeduplicationWorkflow

# Default: uses vLLM with google/embeddinggemma-300m
workflow = TextSemanticDeduplicationWorkflow(
    input_path="input_data/",
    output_path="./results",
    cache_path="./sem_cache",
    n_clusters=100,
    eps=0.07,
    perform_removal=True
)
result = workflow.run()

Checklist

• I am familiar with the Contributing Guide.
• New or Existing tests cover these changes.
• The documentation is up to date with these changes.

[greptile-apps]

Unresolved git merge conflict markers

Lines 117–120 contain raw <<<<<<< HEAD, =======, and >>>>>>> upstream/26.04-staging markers. These were not resolved before committing and will cause the Fern documentation build to fail or render the conflict markers verbatim to users.

The upstream side of the conflict contains a useful comment that should be retained:

Suggested change

	<<<<<<< HEAD
	=======
	# result.metadata contains: total_time, num_duplicates, kmeans_time, pairwise_time
	>>>>>>> upstream/26.04-staging
	# result.metadata contains: total_time, num_duplicates, kmeans_time, pairwise_time