Host allocation sharing via set final data - construction->destruction

[mkinsner]

I'm not completely sure of this fix, but I think "destruction" fits better here since set_final_data is called post-construction. Maybe better to say "valid at buffer destruction" instead of "after", but that matters less.

[mkinsner]

The existing wording seems to reflect one very specific use case of set_final_data, where an initial host allocation which was bound to the buffer becomes unavailable and needs to be replaced with an alternate host allocation, during the buffer's lifetime. There are however other use cases for set_final_data, and the pointer set through that API only matters across the buffer destruction boundary AFAIK.

[gmlueck]

Note that the paragraph you changed will be removed in #298 along with the entire section that contains it. Change #298 is currently pending agreement on the exact semantics of set_final_data.

I'm not really opposed to the change you propose, but I think we need a bigger clarification to this part of the spec.

[keryell]

Let's commit this anyway since even if it is to be replaced later, it is still an improvement today.

[gmlueck]

This PR needs one more reviewer.

[gmlueck]

We could use one more reviewer here before today's WG meeting.

[AerialMantis]

Good catch, looks good.

[gmlueck]

WG approved