fix Qwen3Embed panic when concurrency

[st0nie]

Original code get the rwlock only when embedding is done

When we have multiple concurrent embedding at the same time, kvcache from one request may pollute another embedding process

Like:

Failed to process embedding for "/zion0/smbtest/1111 (1).txt":
embed_file panicked: called `Result::unwrap()` on an `Err` value: shape
mismatch in broadcast_add, lhs: [10, 16, 335, 654], rhs: [10, 1, 335,
335]

Now we hold the write lock during the whole procedure, allow preprocessing to be performed concurrently; but vector generation will not be concurrent.