mtl_tts: Add device/map_location args

[akx]

This makes example_tts.py work OOTB on my Macbook – otherwise it would try to load tensors onto CUDA devices, which are notably missing.

Should also make CPU inference work similarly better.