feat: concurrent embedding, GitHub ZIP download, read offset/limit

[yangxinxin-7]

Performance improvements:

• upload_directory: three-phase approach (collect → pre-create dirs →
  concurrent upload via asyncio.Semaphore, limit 8). Memoized mkdir
  eliminates redundant AGFS calls.
• tree_builder: replace recursive file-by-file move with single agfs.mv()
  wrapped in asyncio.to_thread
• TextEmbeddingHandler: offload blocking embed() to thread pool
• EmbeddingQueue: configurable max_concurrent workers via
  EmbeddingConfig.max_concurrent (default 1)

New features:

• CodeRepositoryParser: use GitHub archive ZIP API instead of git clone
  for GitHub URLs without a specific commit (faster, no git history).
  Includes Zip Slip validation.
• read()/read_file(): add offset/limit line-slicing, propagated through
  all client/service/HTTP/CLI layers

[CLAassistant]

All committers have signed the CLA.