fix(dcgm): harden Init/Shutdown refcount misuse

[rvatkar]

Summary

• Fix dcgm.Init so negative init-counter state returns nil, err without mutating the counter or returning cleanup.
• Fix dcgm.Shutdown so calls before init return an error without decrementing the counter.
• Add regression coverage for invalid shutdowns, negative init counter state, cleanup after manual shutdown, and concurrent invalid shutdown calls.

Testing

• go test ./pkg/dcgm -run 'Test(InitWithNegativeCounterReturnsErrorWithoutCleanup|ShutdownBeforeInitDoesNotCorruptCounter|CleanupAfterManualShutdownDoesNotCorruptCounter|
  ConcurrentShutdownBeforeInitDoesNotCorruptCounter)' -count=1 -v
• go test ./pkg/dcgm -count=1 -v
• go test ./tests -run TestDeviceCount -count=1 -v
• go test -race ./tests -run TestDeviceCount -count=1 -v
• go test -race ./pkg/dcgm -count=1 -v