Add `QTensor::quantize_onto` to remove a redundant dtoh copy?

[EricLBuehler]

Currently, QTensor::quantize:

• Take a tensor, assume it is on the GPU for this example
• Copies the data to the CPU
• Quantizes on the CPU
• Copies the data back from the CPU to the GPU

In particular, this is an unnecessary copy (2 copies total) if the tensors are already on the CPU. Perhaps a QTensor::quantize_onto function would be better, as it would:

• Take a CPU tensor
• Quantize on the CPU
• Copy the data to the GPU

This means there is only one copy. I have implemented this here: EricLBuehler/candle#12, I would appreciate any thoughts on whether this would be a good addition here.