flux_respond_error: accept an errnum of 0

[SteVwonder]

It was discussed on Slack that flux_respond_error considers errnum == 0 to be invalid and errors out with EINVAL. From the man page:

flux_respond_error() returns an error response to the sender. errnum must be non-zero. If errmsg is non-NULL, an error string payload is included in the response. The error string may be used to provide a more detailed error message than can be conveyed via errnum.

This was surprising to both @grondo and I. One suggestion was to allow an errnum of 0 for flux_respond_error since we've been shifting to using descriptive errmsgs anyways. Maybe we only allow an errnum of 0 when an errmsg is provided?

[garlick]

I agree this is a problem area. It's easy to encounter an error and call flux_respond_error (...errno) but find that errno wasn't actually set by the function that failed. No response is sent so someplace else hangs waiting for it. At best the flux_respond_error() failure is logged. Meanwhile stuff is hung and you don't know why.

Since RFC 3 requires the errnum field in the response to be nonzero on error, we should maybe just substitute a generic errno if it is called with errnum = 0?