Sentence Transformers CrossEncoder support

Summary

This PR adds model-side support for loading the reranker through the modular
Sentence Transformers CrossEncoder API:

import torch
from sentence_transformers import CrossEncoder

model = CrossEncoder(
    "KaLM-Embedding/KaLM-Reranker-V1-Large",
    trust_remote_code=True,
    device="cuda",
    model_kwargs={"dtype": torch.bfloat16, "chunk_size": 4},
)

scores = model.predict([(query, document)])
rankings = model.rank(query, documents)

The document is encoded on the encoder side, while the task instruction and
query remain on the decoder side. The default Sigmoid activation returns
P(yes); passing activation_fn=torch.nn.Identity() returns
yes_logit - no_logit.

chunk_size=N performs mask-aware mean pooling over each consecutive group of
N encoder token states. Set chunk_size=None to disable encoder compression.

Files

The PR changes only these six adapter/configuration files:

• kalm_reranker.py
• kalm_reranker_utils.py
• kalm_cross_encoder.py
• kalm_cross_encoder_config.json
• modules.json
• config_sentence_transformers.json

It does not modify or upload the repository README, model weights, tokenizer
artifacts, generation configuration, or Transformers config.json.

Compatibility

• sentence-transformers>=5.6,<6
• transformers>=5.3,<6
• PyTorch inference backend only

The adapter code is shared with the validated
KaLM-Embedding/KaLM-Reranker-V1-Nano CrossEncoder integration. Small and
Large retain their existing model-specific Transformers configurations and
weights.