gpu: ion: optimize zero'ing of pages

Instead of using __GFP_ZERO, vmap high-order pages for zero'ing all at
once (or in a few large chunks).

Test results show a ~25% speedup for allocations that are fulfilled from
lowmem, and ~100% speedup for allocations that are fulfilled from
highmem.

Change-Id: I57270361ff195780370e1c6f458985bdbb641efe
Signed-off-by: Mitchel Humpherys <mitchelh@codeaurora.org>