[libc] Improve copying system vectors to the GPU

[libc] Improve copying system vectors to the GPU

Summary:
This implementation was buggy and inefficient. Fine-grained memory can
only be allocated on a page-level granularity. Which means that each
allocated string used about 4096 bytes. This is wasteful in general, and
also allowed for buggy behaviour. The previous copying of the
environment vector only worked because the large buffer size meant that
we would typically have a null byte after the allocated memory. However
this would break if the vector was larger than a page. This patch
allocates everything into a single buffer. It makes it easier to free,
use, and it more correct.