feat(connector): support multi-PegaServer for cross-node TP

[xiaguan]

Summary

• Support vLLM TP16 across 2 nodes where each node runs its own PegaServer (each managing 8 GPUs)
• Add PEGAFLOW_ENDPOINTS env var / pegaflow.endpoints extra_config for multi-server endpoint resolution
• Connector maps global ranks to local (server-scoped) tp_rank/tp_size/world_size — each PegaServer sees itself as TP8 transparently
• Scheduler queries all servers, takes min(hit_blocks), unpins excess blocks on over-hitting servers
• Single-server (num_servers=1) behavior is fully backward compatible, zero overhead

Changed files

File	Change
python/pegaflow/connector/common.py	Add engine_clients field to ConnectorContext
python/pegaflow/connector/__init__.py	Multi-endpoint resolution + local rank mapping
python/pegaflow/connector/scheduler.py	Multi-server query with min-reduce + unpin excess

Design

• No Rust changes, no vLLM changes — purely Python connector layer
• Namespace uses global tp_size for consistent hashing across all ranks
• Worker auto-connects to its local server: server_index = global_rank // local_tp_size
• Scheduler coordinates queries across all servers with proper pin cleanup in all error paths

Test plan

• Single-server config: verify identical behavior (regression)
• TP16 / 2-server: verify correct rank mapping (rank 0-7 → server 0, rank 8-15 → server 1)
• Multi-server cache hit alignment: verify min-reduce and excess unpin
• Error paths: verify no pin leaks on service error / business error / loading state

🤖 Generated with Claude Code