RFC: Change Process to use directory fd

[ghost]

All data accessors now use openat and readlinkat when reading procfs entries, to protect against PID reuse. Also, iterators are now used instead of Vec in any place where a file descriptor is held open. This is an API break, see examples/netstat.rs and examples/process_hierarchy.rs for the most obvious effects on library consumers. I also removed the stat field as suggested in #69. I haven't made any documentation changes yet, but I can if the approach here looks good. I used a few functions from the nix library for convenience, but those are simple enough that they can potentially be re-implemented here if the dependency isn't desired.

A downside of this is that simple cases using stat are now slower. If all you need during iteration is the information from /proc/<pid>/stat, that now takes two additional syscalls to open and close the directory fd. We could speed that up by adding another function to iterate over the pids, and then adding more constructors to the relevant structures (Stat, MemoryMap, Status...) that directly read from /proc/<pid>/<name>. It would have to be noted in the documentation that it's only safe to use one of those constructors while iterating, because of PID recycling; for cases where two or more files need to be read from within an individual /proc/<pid> directory, the fd-holding iterator from all_processes() must be used.

On second thought, a better way to handle those types of iterations would be to put the pid in a separate Pid structure that then gets consumed by a move into a new Stat constructor. That way it can't accidentally be used twice.

If this all sounds good, I'd like to use this as a base to add pidfd support as well, to support race-free waiting and signaling on procfs processes.

Fixes #125

[eminence]

I'm sorry I haven't gotten a chance to look at this yet. I'm going to try to take a look within a week.

[ghost]

No rush. 😃 This is still a very rough draft.

[sunfishcode]

This patch was incorporated into #171, which is now merged.