Avoid draining fresh virtio-serial agent handshakes

[motatoes]

Summary

• stop discarding pending virtio-serial bytes immediately before accepting the next agent gRPC connection
• avoids dropping the fresh HTTP/2 client preface after QEMU golden restore

Validation

• Reproduced failure with drain enabled on dev AWS QEMU golden restore: VM resumed in ~336ms, then agent ping timed out after 30s.
• Removed only the pre-accept drain call, rebuilt golden, and sandbox creation hot-booted successfully.
• Re-enabled drain only, rebuilt golden, and reproduced the 30s timeout again.
• Restored no-drain build, rebuilt golden, and verified sandbox creation succeeds again.
• Ran: GOOS=linux GOARCH=amd64 go test ./cmd/agent