GPU: support for reusing pre-recorded command buffers

The existing back-ends are not designed to support pre-recorded command-buffer reuse.

In such cases, when the command buffer is recorded, the instrumentation/annotation will obtain a query-id to each command. Query ids are provided incrementally or circularly, so submitting the same pre-recorded command buffer will cause the same pre-obtained query id timestamp slot to be clobbered with other commands that have been issued later.

I'm not sure about the best way to address this. The cleanest solution I have in mind would be to maintain a secondary pool of timestamps for these "persistent" queries. We would need new instrumentation/annotation directives too. And it would still not be fail-proof, as the same command buffer could be submitted multiple times in a row.