Fix additional breakage in covering-index patch.

Fix additional breakage in covering-index patch.

CheckIndexCompatible() misused ComputeIndexAttrs() by not bothering
to fill ii_NumIndexAttrs and ii_NumIndexKeyAttrs in the passed
IndexInfo.  Omission of ii_NumIndexAttrs was previously unimportant,
but now this matters because ComputeIndexAttrs depends on
ii_NumIndexKeyAttrs to decide how many columns it needs to report on.

(BTW, the fact that this oversight wasn't detected earlier implies
that we have no regression test verifying whether CheckIndexCompatible
ever succeeds.  Bad dog.  Not the job of this patch to fix it, though.)

Also, change the API of ComputeIndexAttrs so that it fills the opclass
output array for all column positions, as it does for the options output
array; positions for non-key index columns are filled with zeroes.
This isn't directly fixing any bug, but it seems like a good idea.

Per valgrind failure reports from buildfarm.

Alexander Korotkov, tweaked a bit by me

Discussion: https://postgr.es/m/CAPpHfduWrysrT-qAhn+3Ea5+Mg6Vhc-oA6o2Z-hRCPRdvf3tiw@mail.gmail.com