[benchmarks] reduce setup time for RandomMemcpy benchmark

Allocating the memory buffer for RandomMemcpy can be very slow. In
some manual testing on Eve, the buffer allocation appeared to take 4-8
seconds.

This is with a debug build which uses -O0, and the std::vector
constructor will produce code that does a function call for every byte
in the 128MB buffer.  (In release builds, the compiler will optimise
that to a memcpy().  When unoptimised, it's very slow.)

Optimize the buffer allocation, by using a primitive array, rather
than a std::vector. We fill the array before reading from it, to
ensure that we do not cause any reads of unitialized memory.

The net result of the optimization is that the time to run the
benchmarks in "unit-test-mode", on my Eve, drops from ~2 minutes to
~30 seconds.

Bug: FLK-124 #done
Test: /pkgfs/packages/zircon_benchmarks/0/test/zircon_benchmarks (on eve)
Test: /pkgfs/packages/zircon_benchmarks/0/test/zircon_benchmarks -p (on eve)
Change-Id: I449a304a60ccee6a137e9371ca4455b8840fedde