fix(embedding): set kv_unified=True when embedding=True to enable batch processing

[SanjanaB123]

Fixes #2216

Problem

When using embedding=True, n_seq_max is set to 256 but kv_unified
remains False (the default). This causes n_ctx_seq to be computed
as n_ctx / n_seq_max = 256 tokens per sequence, making batch
embeddings essentially unusable for inputs longer than 256 tokens.

Fix

Set self.context_params.kv_unified = True inside the if embedding:
block, which allows each sequence to use the full context window
instead of n_ctx / n_seq_max.

Testing

Verified that both single and batch embeddings work correctly after
the fix:

• Single embedding: ✅ works, returns 384-dim vector
• Batch embedding with long inputs: ✅ works, returns correct embeddings