refactor(server): use ComputeDriver RPC surface in-process

[drew]

Summary

Bind openshell-server directly to the generated ComputeDriver RPC surface in-process so the server and driver share one method surface without network transport. Keep public GetSandbox and ListSandboxes store-backed, and reconcile that store from driver snapshots without regressing newer watch updates or pruning live sandboxes.

Related Issue

N/A

Changes

• Remove the local ComputeBackend adapter and call ComputeDriver RPC handlers directly with tonic::Request.
• Keep public sandbox get and list store-backed while reconciling the store from ListSandboxes and GetSandbox.
• Fix reconcile races so stale list snapshots cannot overwrite newer watch state or delete a sandbox that now exists in the driver.
• Preserve driver gRPC status mapping for already-exists and precondition errors, and update the gateway architecture doc.

Testing

• mise run pre-commit passes
• Unit tests added/updated
• E2E tests added/updated (not applicable)
• cargo test -p openshell-server compute:: -- --nocapture
• cargo test -p openshell-driver-kubernetes grpc::tests -- --nocapture
• mise run pre-commit failed in an unrelated openshell-cli --test sandbox_create_lifecycle_integration run because port 8080 was already in use.

Checklist

• Follows Conventional Commits
• Commits are signed off (DCO)
• Architecture docs updated (if applicable)

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[maxamillion]

LGTM

+1 to the convergence of the api call regardless of local or remote gRPC