feat: Add request cancellation to C++ gRPC client

[yinggeh]

Summary

Adds per-request cancellation to the C++ gRPC client, mirroring the existing Python tritonclient.grpc cancellation interfaces.

• AsyncInfer gains an optional trailing CallContext** ctx_out. On Error::Success, the caller receives a heap-allocated CallContext whose Cancel() method calls grpc::ClientContext::TryCancel(). Cancel() after natural completion is a safe no-op.
• AsyncInferMulti gains an optional trailing std::vector<CallContext*>* ctxs_out — one handle per fanned-out request. Cancellation is per-request; the multi callback still fires exactly once after every leaf produces a result.
• StopStream gains a bool cancel_requests = false. When true, the streaming RPC is TryCancel'd and the stream callback receives one final InferResult whose status contains Locally cancelled by application!.
• After a cancelled stream, StartStream can be called again — grpc_context_ is rebuilt in place because grpc::ClientContext is non-movable and a cancelled context cannot be reused.
• All cancellation paths surface the same Python-parity message Locally cancelled by application!, matching tritonclient.grpc._utils.get_cancelled_error() in the Python client.

Backwards compatibility

Every new parameter defaults to nullptr / false. Existing call sites compile and behave exactly as before; cancellation is strictly opt-in.

Testing

New src/c++/tests/grpc_cancellation_test.cc (gtest) with 7 cases:

Test	Covers
TestGrpcAsyncInfer	Python test_grpc_async_infer parity (1:1)
TestGrpcAsyncInferCancelAfterCompletionIsNoOp	Cancel-after-finish is safe, no double callback
TestGrpcAsyncInferWithoutContextStillCompletes	Default arg path is unchanged
TestGrpcAsyncInferMulti	Cancels requests 0 and 2, lets request 1 complete; verifies per-request cancel + result-order preservation + single multi-callback fire
TestGrpcStreamInfer	Python test_grpc_stream_infer parity (1:1)
TestGrpcStreamCancelWithoutInfer	Cancel an empty stream still emits the cancel message
TestGrpcStreamCancelThenRestart	Cancelled stream → fresh stream → successful inference

Related

triton-inference-server/server#8775

[whoisj]

LGTM

[mudit-eng]

Pointer to pointer is tricky and error prone. Why not pass as a reference?

[yinggeh]

Because the argument is optional. A reference cannot be null.

[mudit-eng]

Help me understand. Is it ctxs_out that would be null or the vector entries inside it? Asking because can we have an empty vector?

[yinggeh]

The pointer value doesn't matter and can be null. See example in the document
https://github.com/triton-inference-server/client/pull/896/changes#diff-b335630551682c19a781afebcf4d07bf978fb1f8ac04c6bf87428ed5106870f5R603-R616