[BugFix] Fix new nightly failures

[LucasWilkinson]

Fix nightly failures:

tests/models/language/pooling_mteb_test/test_jina.py::test_embed_models_mteb[model_info0]
tests/models/language/pooling/test_token_classification.py::test_modernbert_models[float-disham993/electrical-ner-ModernBERT-base]
tests/models/language/pooling_mteb_test/test_st_projector.py::test_embed_models_mteb[model_info1]
tests/models/multimodal/pooling/test_siglip.py::test_models_text[float-google/siglip-base-patch16-224]
tests/models/multimodal/pooling/test_siglip.py::test_models_text[float-google/siglip2-base-patch16-224]
tests/models/multimodal/pooling/test_siglip.py::test_models_text_image_no_crash[float-google/siglip-base-patch16-224]
tests/models/multimodal/pooling/test_siglip.py::test_models_text_image_no_crash[float-google/siglip2-base-patch16-224]

examples/offline_inference/spec_decode.py --test --method eagle --num_spec_tokens 3 --dataset-name hf --dataset-path philschmid/mt-bench --num-prompts 80 --temp 0 --top-p 1.0 --top-k -1 --tp 1 --enable-chunked-prefill --max-model-len 2048

Caused by #28579

All tests above pass; running ready-run-all-tests

[gemini-code-assist]

Code Review

This pull request addresses several nightly failures by correcting the shape of a dummy block table and introducing an unpadded method to handle attention metadata for speculative decoding. The block table shape fix is a critical correction. I've also suggested an improvement to the unpadded method to ensure metadata correctness by recomputing max_query_len and max_seq_len for the unpadded data, which is important for correctness and performance.

[DarkLight1337]

The doc build failure is unrelated and fixed by #29619