Add sparse text and document embedders to the vLLM integration #4160

Is your feature request related to a problem? Please describe.
The vLLM integration only provides dense embedders. Using vLLM-hosted sparse models requires custom components.

Describe the solution you'd like
Add:

• VLLMSparseTextEmbedder: returns a SparseEmbedding for a string.
• VLLMSparseDocumentEmbedder: populates each document’s sparse_embedding field.

Support sync and async execution, with configuration consistent with the existing vLLM embedders.

Describe alternatives you've considered
Custom components or existing sparse embedders that require a different inference backend.

Additional context
N/A