`ModernBertEmbedder` hardcodes mean pooling; ignores `1_Pooling/config.json`

[amgix-dev]

Problem

ModernBertEmbedder always uses mean pooling, regardless of the model's Hugging Face config.

In rust/src/embeddings/local/modernbert.rs, pooling is hardcoded at load time:

Ok(ModernBertEmbedder {
    model,
    tokenizer,
    device,
    pooling: Pooling::Mean,  // always mean
})

Many ModernBERT embedding models ship with 1_Pooling/config.json specifying CLS pooling, for example:

• ibm-granite/granite-embedding-97m-multilingual-r2 → pooling_mode_cls_token: true
• Alibaba-NLP/gte-modernbert-base → pooling_mode_cls_token: true

Using ModernBertEmbedder::embed() on these models ignores the model's declared pooling strategy.

Expected behavior

On load, resolve pooling from the model repo:

1. If 1_Pooling/config.json exists, use it (pooling_mode_cls_token, pooling_mode_mean_tokens, or pooling_mode)
2. Else fall back to config.json → classifier_pooling ("cls" / "mean")
3. Else default to mean

BertEmbedder already accepts pooling: Option<Pooling> in new(); ModernBertEmbedder should follow the same pattern.

Suggested fix

• Add pooling: Option<Pooling> to ModernBertEmbedder::new
• Add a small helper to load pooling from the HF repo (1_Pooling/config.json + fallback)
• Pass pooling through from TextEmbedder::from_pretrained_hf (currently ignored for ModernBERT)

Impact

Any user calling ModernBertEmbedder::embed() on CLS-pooled models gets incorrect embeddings.

[Ha1baraA11]

Hi, I'd like to work on this issue.

I plan to update ModernBertEmbedder to honor an explicit pooling choice and resolve the repository metadata fallback described here, while preserving mean pooling as the default. I'll also add focused tests for CLS, mean, and fallback behavior.

Could a maintainer assign this issue to me before I begin?