When using a read-only CPU mapping, we may encounter stale buffer
contents.  For example, the Piglit primitive-restart test offers the
following scenario:

   1. Read data via a CPU map.
   2. Destroy that buffer.
   3. Create a new buffer - obtaining the same one via the BO cache.
   4. Call BufferSubData, which does a GTT map with MAP_WRITE | MAP_ASYNC.
      (We avoid set_domain for async mappings, so no flushing occurs.)
   5. Read data via a CPU map.
      (Without explicit clflushing, this will contain data from step 1!)

Otherwise, everything ought to work, keeping in mind that we never use
CPU maps for writing - just read-only CPU maps.

This restores the performance gains after Matt's revert in commit
71651b31

.

v2: Do the invalidate later, and even when asking for a brand new map.
v3: Add more comments from Chris.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>