further clarify a clCreateBuffer with SVM pointer error condition

[karolherbst]

It should also be invalid to use SVM host_ptr smaller than size in the CL_MEM_COPY_HOST_PTR case

[karolherbst]

Not sure if this has been specifically discussed in the OpenCL WG, but it feels like an oversight as this is almost guaranteed to cause segmentation faults if not caught.

[karolherbst]

Though I think if runtimes are supposed to check those pointers here, they might as well check if the access is generally out of bounds, e.g. using an SVM allocation + 0x100 of size 0x1000 and then using 0x1000 as the size to clCreateBuffer, which will cause undefined behavior if not checked by the runtime.

(But so does passing random host ptrs with invalid sizes where runtimes can't bound check)

[bashbaug]

Linking with #875, which I believe prompted the current text.

This seems reasonable to me, though if we merge these changes we really ought to have a negative CTS test to check for the new error behavior..

[bashbaug]

Merging as discussed in the August 13th teleconference.