fix(lxdclient): fix context race condition in ProtocolLXD WithContext

[whywaita]

Summary

• Fix race condition where ProtocolLXD.WithContext() was called before APICallMutex.Lock(), causing concurrent goroutines to overwrite each other's context on the shared client
• Fix context not being reset on error paths, leaving the client with a canceled context for all subsequent calls
• Fix setLXDHostResourceCache modifying client context without holding the Mutex

Changes

• server/pkg/api/server_host.go: Move WithContext after APICallMutex.Lock(), add defer reset on all paths
• server/pkg/lxdclient/resource.go:
  • GetResourceFromLXD: Move WithContext after Mutex acquisition + defer reset
  • GetResourceFromLXDWithClient: Remove xsemaphore and internal WithContext/reset (callers now own this responsibility)
• server/pkg/metric/scrape_lxd.go: Add WithContext + defer reset after TryLock
• server/pkg/resourcecache/resource_cache.go: Add APICallMutex.Lock() + WithContext + defer reset
• server/go.mod: Remove xsemaphore dependency
• server/pkg/lxdclient/connect_test.go: Add regression tests for Mutex-based concurrency control

Test Plan

• go test ./... — all tests pass
• go vet ./... — no issues
• go build ./... — build succeeds

Notes

• Root cause fix for production incident where context canceled errors caused no available instance failures despite available capacity
• Unified pattern applied to all call sites: Mutex Lock → WithContext → defer reset → API call