vLLM pooling models documentation

vLLM pooling models docs. Carries the important caveat verbatim: "We currently support pooling models primarily for convenience. This is not guaranteed to provide any performance improvements over using Hugging Face Transformers or Sentence Transformers directly" (tracking issue 21796). Documents the four pooling tasks (classify/embed/token_classify/token_embed), the five pooling types (CLS/LAST/MEAN/ALL/STEP), the cross-encoder / late-interaction / bi-encoder score types, and the supported embedding and cross-encoder architecture tables.