Improve a potential out-of-memory edge case in CowData by propagating errors out of _copy_on_write.

[Ivorforce]

I am not making much progress on #100619, so I'm making another small PR about an inconsistency in the CowData class that I noticed while spending so much quality time with it.

It's an edge case where _copy_on_write fails to allocate a new buffer (out of memory).
Previously, the error was logged, the non-1 refcount was ignored, and the shared data may have been reallocated in later instructions, possibly invalidating it for other owners.
Instead, this PR changes behavior such that an out-of-memory error is recognized and propagated upwards, such that no shared data is invalidated.

This is unlikely to have an effect during most runs, but it may improve error behavior in very rare edge cases.

The return value of _copy_on_write was previously ignored in all other spots, and still is with the new code. In those cases, it is likely that shared data is overwritten as a result (though not completely invalidated as in the resize case).