`Foreign.Marshal.Pool` used to call `malloc` once for each allocation
request. Each `Pool` maintained a list of allocated pointers, and
traverses the list to `free` each one of those pointers. The extra O(n)
overhead is apparently bad for a `Pool` that serves a lot of small
allocation requests.

This patch uses the RTS internal arena to implement `Pool`, with these
benefits:

- Gets rid of the extra O(n) overhead.
- The RTS arena is simply a bump allocator backed by the block
  allocator, each allocation request is likely faster than a libc
  `malloc` call.

Closes #14762 #18338.