TEN audio/video frame should truly leverage video memory, where lock/unlock would be real

[halajohn]

The extra difficulty might lie in how to better ensure compatibility across different inference frameworks while minimizing copies.