Load features onto the GPU in batches to support arbitrarily long audio

[lachesis]

We tested with large model on nvidia 1070 (8GB) with default settings. We also verified that it works with chunking disabled, and with language detection enabled.

• 1 hour of 16-bit stereo PCM audio at 48kHz takes about 600MiB of CPU RAM, so there is no issue with preprocessing all of the audio at once.
• GPU batches default to 2 chunks, but this can be tweaked (and should be for cards with more RAM).
• Language detection only looks at the first GPU batch worth of audio.
• This feature conflicts with translation, so if long audio is submitted and translation is enabled, it will be disabled on the fly and no translation will be done.
• Chunking really doesn't need to be disabled any more, even for relatively low RAM cards. We made the chunking memory threshold configurable in settings.py and defaulted it to 4GB.

Processing all of "12 Angry Men" (1 hour 36 minutes) on large model with beam size 5 took 656666ms on a 1070 Ti and the results were reasonable.

Thanks @richardklafter for pair programming this patch.