Tighter typing of Proxy_server.sym_block_caching_time (follow-up from "lib_proxy: don't call Lwt_unix.sleep directly")

The following discussion from !4533 (merged) should be addressed:

• @raphael-proust started a discussion: (+1 comment)

  LGTM too.

  As a nitpick, the change to a record could have been a good time to change the sym_block_caching_time. I'm thinking specifically about:

  • making sure it's a positive value (currently I can't find a check),
  • making it a Ptime.span (although that's a bit more intrusive, especially to do it cleanly right from the argument parsing) (this is to avoid having to remember the unit (seconds) of the value which is only documented in the cmdliner parsing block, not anywhere near the use site).

  That can be for a follow-up MR more probably.