Langroid

- Adds a remote embeddings server for sentence transformer embeds, using grpc for remote procedure calls, allowing running langroid and generating local embeddings on different machines; currently, this functionality is not handled by tools such as ollama and litellm.

- If possible, uses data parallelism for local embeds; adds configurable batch sizes for all embeddings.

- Additionally, improves imports for parsing and vector stores.

run_batch_tasks` now takes a `turns` param to limit number of turns in each task (default = -1 meaning infinite)

ImagePdfParser`: extract text from image-based pdfs. This means `DocChatAgent` now works with image-based pdfs.

LanceDB: when creating schema, sort fields to prevent variable ordering and consequent mismatch between stored table and new docs during ingest.

LanceDB: allow adding docs with new metadata fields that were not in original config.document_class

Fix regex in Relevance Extractor, used in the [`extract_numbered_segments` function](https://github.com/langroid/langroid/blob/main/langroid/parsing/utils.py).
This was causing the segments to only include portions before a newline,
which meant that for certain documents, the extraction was losing big portions of the text between
sentence/segment markers. This fix should vastly improve relevance extractor and get rid of issues
seen when using this in Doc Chat.