fix(serialization): Slice 1-D multibyte data as bytes for pwrite

[Eta0]

1-D Multibyte Data Slicing

This change fixes a bug in TensorSerializer's write routine that would incorrectly slice 1-D multibyte data using its original dtype width rather than by individual bytes. The effect of this was that if attempting to write a buffer only wrote part of it, slicing the beginning of the buffer away to then attempt to write the rest of it would create an invalid slice, and then fail the write routine with an exception.

For example, if writing a 32-byte buffer of 4x float64 elements partially succeeded and wrote 12 bytes, it would previously essentially attempt to then continue by writing buffer[12:], but mistakenly skip ahead by the width of 12x float64 elements (i.e. 96 bytes) instead of 12 bytes. It would then find that it had no more data to write, and then check that it wrote everything, and fail. To fix this, it now casts buffers to an unsigned byte type before slicing if it is not one already.

This only affects 1-D data because multidimensional arrays passed to TensorSerializer's write routine were already correctly converted to 1-D arrays of bytes. Only ones that were already 1-D, but the wrong width, were missed.

Code version update to v2.9.1

This change additionally updates the code version from v2.9.0 to v2.9.1 so that this fix can be published as a bugfix release.

[wbrown]

I have questions!