Feature: Add an Exec channel to run commands inside a running actor (E2B-style sandbox exec)

**Summary**

Add a first-class streaming Exec RPC that lets a client run an arbitrary command inside a running actor's gVisor sandbox, with bidirectional stdin / stdout / stderr, exit code, optional TTY, and window resize — analogous to
kubectl exec for Pods, or to E2B's Sandbox.commands.run / Sandbox.pty.create APIs for sandboxed code-execution platforms.

This is a management-plane channel, separate from the L7 actor traffic that atenet routes. It does not change the Actor lifecycle and is purely additive.

**Motivation**

Substrate is positioned as the runtime for agentic / sandboxed code-execution workloads (see README.md:199 "Sandbox Demo" and docs/api-guide.md:168 LangChain section). For that class of workload, "run a command in the
sandbox and stream the result back" is the single most fundamental primitive — it's what platforms like E2B, Modal sandboxes, and Daytona expose to end users. Today Substrate has no equivalent:

- pkg/proto/ateapipb/ateapi.proto — Control service exposes Create / Suspend / Resume / Delete / Get / List only.
- internal/proto/ateletpb/atelet.proto — AteomHerder exposes Run / Checkpoint / Restore only, all unary, Streams: [] confirmed in the generated atelet_grpc.pb.go:221.
- internal/proto/ateompb/ateom.proto — Ateom mirrors the same three unary RPCs.
- cmd/ateom-gvisor/runsc.go wraps runsc create / start / wait / kill / checkpoint / restore — runsc exec is not wrapped, even though it exists upstream.
- cmd/kubectl-ate/internal/cmd/ has create / get / delete / suspend / resume / logs / admin — no exec. logs_actors.go:73 falls back to kubelet Pods().GetLogs(...), i.e. log streaming is delegated to Kubernetes today.

Concretely this blocks the following use cases:

1. Agent tool calls that exec untrusted code (the canonical E2B use case). Today an agent has to embed an HTTP/RPC server in its image and let atenet route to it — every template author re-implements the same exec-server
boilerplate.
2. Interactive debugging of a stuck actor. SREs currently have to kubectl exec into the privileged atelet DaemonSet pod (which mounts the host's /run/ateom-gvisor) and hand-craft a runsc -root
/run/ateom-gvisor/actors/::/runsc-state exec ... invocation. This works but bypasses the control plane entirely (no audit, no lock against concurrent suspend/checkpoint, RBAC granularity = "all actors on the
cluster").
3. One-shot setup commands for templates (npm install, db migrate, …) without baking them into the container image or the workload entrypoint.