When converting a vertex buffer by using GPU, the conversion buffer
doesn't need to be host mappable. Hence the conversion buffer can
be allocated on device local memory for faster GPU access times.

Bug: angleproject:3534
Change-Id: I2efabec20186992479920bddd3abd36f9c13babc
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2108706
Commit-Queue: Mohan Maiya <m.maiya@samsung.com>
Reviewed-by: Jamie Madill <jmadill@chromium.org>