Optimize work retrieval for RenderEngineThreaded

* Avoid copying the work functor when retrieving the next task to run
* Run the work functor in its own scope, so that it's destroyed before RenderEngine waits for more work

These optimizations become more significant since each work functor can have lambda captures, which includes smart pointers and gralloc buffers. The latter is significant, as if the work functor is persisted for too long, we retain gralloc memory for longer than intended.

Bug: 378206452
Flag: EXEMPT BUGFIX
Test: builds
Test: Transient dmabuf overhead no longer present
Test: courage
Change-Id: Icfc25443c767072376c5aed4d725a33b4e68490c