The meaning of the limit changed in 7.4, it now points one past the
end. Adjust code accordingly.

The php_stream_read() and php_stream_write() functions now return
an ssize_t value, with negative results indicating failure. Functions
like fread() and fwrite() will return false in that case.

As a special case, EWOULDBLOCK and EAGAIN on non-blocking streams
should not be regarded as error conditions, and be reported as
successful zero-length reads/writes instead. The handling of EINTR
remains unclear and is internally inconsistent (e.g. some code-paths
will automatically retry on EINTR, while some won't).

I'm landing this now to make sure the stream wrapper ops API changes
make it into 7.4 -- however, if the user-facing changes turn out to
be problematic we have the option of clamping negative returns to
zero in php_stream_read() and php_stream_write() to restore the
old behavior in a relatively non-intrusive manner.

Instead of guessing what the address of the first arena allocation
is going to be, embed the sentinel in the oplog_list structure
directly.

In particular, make sure that everything using zmm is released
before zmm is shut down. phpdbg currently gets away with this,
because either a) its custom handlers are used and no auto-free
happens or b) the system allocator is used and no auto-free happens.
With the tracking allocator for asan this no longer works.

Uses system allocator when backing up settings ... let's avoid
unnecessary confusion.

We want the class declaration to be available while compiling the
parent class.