K8s deployment fixes

[Bellk17]

Motivation

Fix several issues in the K8s deployment manifests and Dockerfile that prevented successful deployment to a production RKE2 cluster with AMD MI325X GPUs.

Technical Details

• Fix Dockerfile build: Bumped rust:1.82 to rust:latest (edition 2024 required by redb crate) and added protobuf-compiler + libprotobuf-dev to the builder stage
• Fix configmap config syntax: Changed [partitions.default] (TOML map) to [[partitions]] with name = "default" (TOML array-of-tables) to match SlurmConfig parser; replaced [cluster] name = ... with top-level cluster_name = ...; added required nodes field
• Fix operator args ordering: Removed run subcommand from args — --controller-addr and other flags are top-level clap args that must appear before any subcommand, and run is already the default
• Simplify spool volume: Replaced volumeClaimTemplate with emptyDir for the spurctld StatefulSet to avoid PVC provisioning issues on clusters with certain storage backends (e.g., Longhorn)
• Fix spurrestd controller address: Prepended http:// to the --controller arg in spurrestd.yaml — spurrestd passes the address directly to tonic's connect() which requires a scheme

Test Plan

• docker build -f deploy/Dockerfile completes successfully
• All manifests apply cleanly to an RKE2 cluster
• spurctld starts and runs the scheduler loop
• spur-k8s-operator connects to spurctld and registers GPU nodes
• spurrestd connects to spurctld without 500 errors
• spur nodes shows registered MI325X node with correct resources

Test Result

Deployed to a single-node RKE2 cluster with AMD MI325X GPUs. All Spur control plane components (spurctld, spur-k8s-operator, spurrestd) start and communicate successfully. GPU node registered with correct CPU, memory, and GPU resources.