Add a client for Gazelle

[handrew]

Created an abstraction over some of the code in the example notebooks (infer, infer-quantized) so that the user can simply:

from gazelle import GazelleClient
client = GazelleClient(quantization="8-bit")
resp = client.infer(audio, prompt="What does the following audio say? \n <|audio|>")
print(resp)

Because of limited resources, I have not yet been able to extensively test it. It seems to work regularly with quantization, but I occasionally get some errors w/ the conv layer without quantization.

Would be happy to take feedback!