sort: Avoid linear realloc loops

[markjdb]

These interact quite poorly with MRS with revocation enabled, which implements realloc() by always creating a new allocation. Even a modification to return the original pointer if its bounds satisfy the request doesn't help much there.

Apply a simple bandaid: grow the allocation exponentially. This has no major downside except when virtual address space is scarce, and that's something we don't typically worry about. There doesn't appear to be any explicit justification for growing buffers so conservatively.

This yields a significant reduction in the amount of time needed to build a cscope database (cscope invokes sort(1) under the hood).

[bsdjhb]

This may be a quirk worth documenting in the CHERI C programmers guide. I feel like I've seen various programs that cap geometric growth so that at some point only a fixed count of items is added instead of always doubling.

[markjdb]

This may be a quirk worth documenting in the CHERI C programmers guide. I feel like I've seen various programs that cap geometric growth so that at some point only a fixed count of items is added instead of always doubling.

See my comment in PR #2133 - I think we can alleviate the problem if jemalloc is willing to loosen its bounds to those of the underlying allocation size, and that's probably closer the right solution.