Compatibility: add `EMBEDDING_SEND_DIM` / `EMBEDDING_TOKEN_LIMIT` and Cohere rerank chunking options

[joonsoome]

LightRAG’s latest update (+1.4.9.9) introduced new embedding and rerank configuration options. We should verify embed-rerank compatibility and add support where missing so LightRAG integrations keep working.

New options to cover:

• EMBEDDING_SEND_DIM: control whether to send the dimensions parameter for OpenAI/Gemini embeddings (Gemini requires true).
• EMBEDDING_TOKEN_LIMIT: token cap used for automatic embedding truncation.
• Cohere rerank chunking:
  RERANK_ENABLE_CHUNKING
RERANK_MAX_TOKENS_PER_DOC

Tasks:

• Audit current embedding/rerank config and request handling for these options (or equivalent aliases).
• If missing, add settings/env aliases and wire them into:
  • Embedding request handling (dimensions on/off + default token limit).
  • Cohere rerank handling (optional chunking by token limit, deterministic aggregation back to doc-level scores).
• Update README/docs with new configuration keys and defaults.
• Add a minimal test or repro checklist for these options.

Acceptance criteria:

• LightRAG can set these env vars without errors and with expected behavior.
• Default behavior is unchanged when the options are unset.
• Documentation is updated with clear examples.