Endianness is determined by host system, not backend

[agoose77]

Version of Awkward Array

main

Description and code to reproduce

We recently made endianness explicit in Awkward Array, with to_buffers producing little-endian buffers, and from_buffers converting to the system byteorder.

This poses a problem on big-endian host systems that use e.g. CuPy, as CUDA devices are little-endian.

Really, endianness is a device property. If we impose the constraints that

• all buffers must belong to the same backend X
• all buffers operated upon by X must have the same device endianness

then we can fix this by taking the desired endianness from the backend.

[jpivarski]

I remember discussing this in a meeting and going down the rabbit hole of checking GPU architectures for endianness. It doesn't look like any GPUs are big-endian, so this is not a high priority. It would be in principle the right thing to do, though.