`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.