BaseIOStream.write(): support typed memoryviews

[madsbk]

This PR implements support of memoryviews with an item size greater than 1 byte.

Currently, Tornado uses len(x) to retrieve the number of bytes of x, which works fine when x is bytes or a memoryview with an item size of 1. However, with an item size greater than 1 we cannot assume that len(x) == x.nbytes.

[madsbk]

Sorry for the close and re-opening of the PR. We have fixed the issue in Dask/Distributed by always casting to bytes before communication: dask/distributed#4555 but I think this PR is still relevant for other downstream projects that might use typed memoryviews.

[bdarnell]

Interesting, I hadn't noticed this aspect of memoryview before (I thought it was just a python-level exposure of the buffer protocol).

It looks like we could alternatively convert the input to a memoryview (unconditionally) and use nbytes instead of len. Is there any particular reason to prefer one form over the other? (I like replacing the union with a concrete type as early as possible, although len() is the natural idiom so choosing a type whose len() is problematic is error-prone).

[madsbk]

It looks like we could alternatively convert the input to a memoryview (unconditionally) and use nbytes instead of len. Is there any particular reason to prefer one form over the other? (I like replacing the union with a concrete type as early as possible, although len() is the natural idiom so choosing a type whose len() is problematic is error-prone).

I think both approaches make sense but if you are not planning to support non-contiguous buffers or fancy data compression, casting memoryviews to cast("B") seems fine than you can use nbytes and len interchangeable.