fix(grpc): commit protobuf-5.x runtime files to source tree (#173)

[waylandzhang]

Summary

Fixes #173 — the gRPC runtime files (*_pb2.py / *_pb2_grpc.py) for the Python llm-service were never committed, so source tags and images built from them were missing the runtime modules and crashed on import. The reported workaround was running make proto-local before building.

Root cause

.gitignore globally ignores **/*_pb2.py and **/*_pb2_grpc.py. The intended unignore for grpc_gen/ was left commented out, so the files stayed untracked. The llm-service Dockerfile copies python/llm-service/ wholesale, so any image CI builds from a tagged source ends up missing the runtime files.

Changes

• Add explicit .gitignore negations for the grpc_gen runtime files
• Force-add the 14 nested protobuf-5.x _pb2.py / _pb2_grpc.py files (matches the protobuf==5.x pin in requirements.txt)
• Regenerate orchestrator stubs to match the committed proto contract (cache-token fields: cache_read_tokens, cache_creation_tokens, cache_creation_1h_tokens — already present in protos/orchestrator/orchestrator.proto)
• Remove stale flat protobuf-6.x duplicates (unimported; the service imports the nested grpc_gen.<pkg> layout, and 6.x-generated code would crash under the pinned 5.x runtime)

Verification

• All staged _pb2.py files report protobuf major version 5
• All grpc_gen modules import cleanly
• Generated code matches the committed .proto contract
• llm-service image build + in-container import smoke confirmed the runtime modules load and cache-token fields exist at runtime

Note for adopters

This fixes the source-build path. The one-command install pulls prebuilt images, so adopters get the fix only once the published llm-service image/tag is rebuilt from a tree containing this commit.