Fix pg_ctl on Windows to reliably detect already-running postmaster.

The previous implementation invoked postgres.exe via cmd.exe, which meant
pg_ctl got the PID of the shell wrapper rather than the actual postmaster
process.  This caused problems in wait_for_postmaster_start(), which tries
to verify that the PID in postmaster.pid matches the one we started.  The
mismatch led to a timing window where pg_ctl could incorrectly report
success when attempting to start an already-running cluster.

Fix by replacing the cmd.exe wrapper with a direct CreateProcess() call.
This gives us the real postgres.exe PID immediately, eliminating the
need for process tree walking or other heuristics to find the actual
postmaster.  We handle I/O redirection using Windows handle-based APIs
instead of shell syntax, which is more reliable anyway.

Author: dbgreen