Pooling::Cls should use first real token (attention mask), not index 0

[amgix-dev]

Problem

Pooling::Cls in rust/src/embeddings/local/pooling.rs always takes the hidden state at sequence index 0:

fn cls(output: &ModelOutput) -> Result<PooledOutputType, anyhow::Error> {
    match output {
        ModelOutput::Tensor(tensor) => tensor
            .get_on_dim(1, 0)  // always position 0
            ...
    }
}

For padded batches, CLS pooling should use the first non-padding token (via attention_mask), not a fixed index 0. This matters when sequences are left-padded.

Pooling::LastToken already has padding awareness in the same file; CLS does not.

Expected behavior

CLS pooling should:

1. Take attention_mask as input (like mean pooling)
2. Use the index of the first real token per sequence
3. Match the behavior used by common ModernBERT embedding checkpoints (e.g. Granite, gte-modernbert)

Impact

CLS-pooled models may produce incorrect embeddings when batching with padding, especially left padding.

Suggested fix

Update Pooling::Cls to accept attention_mask and gather the first real token per row.