Fix unlikely races with GnuTLS, datagrams

Fix unlikely races with GnuTLS, datagrams

Retry some calls if interrupted at inopportune times.
These were found by code inspection.
* src/gnutls.c (gnutls_try_handshake): Simplify by using
new emacs_gnutls_handle_error API.
(emacs_gnutls_write): Remove GNUTLS_E_AGAIN hack since
emacs_gnutls_handle_error now does that.
Use emacs_gnutls_handle_error only on errors.
(emacs_gnutls_read): Retry if gnutls_record_recv returns
GNUTLS_E_INTERRUPTED, to be consistent with emacs_read.
(emacs_gnutls_handle_error): Return 0 on fatal errors,
-1 (setting errno) on ordinary errors, to simplify callers.
Assume that ERR is negative, since it always is now.
Map non-fatal GnuTLS errors to errno values as best we can.
* src/process.c (read_process_output) [DATAGRAM_SOCKETS]:
Retry recvfrom if it is interrupted, to be consistent with
how things are handled when not a datagram channel.
(send_process) [DATAGRAM_SOCEKTS]: If sendto is interrupted,
process pending signals and retry it, to be consistent with
how things are handled when not a datagram channel.