WAR on index mapping when exact and permissive maps differ

[shmsong]

This is a WAR for this upstream issue:

Very counter-intuitively, in ca maps, exact mapping does not imply permissive mapping. The war included in this PR should be a safe change but it isn't a thorough fix of this issue.

The underlying issue is quite similar to this one even though they ended up hitting different asserts.

As shown in the printout from the repro in this PR:

%kernel {
T2_l[ bS17{( 1 * 1 )} ]
   = T0_g[ bS18{( 1 * 1 )} ];
T3_l[ bS16{( 1 * ( 1 * 1 ) )} ]
   = broadcast( T2_l[ bS17{( 1 * 1 )} ] )
T4_g[ iS14{( i4 * ( 1 * 1 ) )} ]
   = T3_l[ bS16{( 1 * ( 1 * 1 ) )} ]
   + T1_g[ iS20{( i4 * ( 1 * 1 ) )} ];

TransformPrinter : 
T0_g[ bS18{( 1 * 1 )} ]
 root domain : (bS0{1},bS1{1})
  Merge: bS0{1} and bS1{1} -> bS18{( 1 * 1 )}
T2_l[ bS17{( 1 * 1 )} ]
 root domain : (bS5{1},bS6{1})
  Merge: bS5{1} and bS6{1} -> bS17{( 1 * 1 )}
T3_l[ bS16{( 1 * ( 1 * 1 ) )} ]
 root domain : (bS7{1},bS8{1},bS9{1})
  Merge: bS8{1} and bS9{1} -> bS15{( 1 * 1 )}
  Merge: bS7{1} and bS15{( 1 * 1 )} -> bS16{( 1 * ( 1 * 1 ) )}
T1_g[ iS20{( i4 * ( 1 * 1 ) )} ]
 root domain : (iS2{i4},bS3{1},bS4{1})
  Merge: bS3{1} and bS4{1} -> bS19{( 1 * 1 )}
  Merge: iS2{i4} and bS19{( 1 * 1 )} -> iS20{( i4 * ( 1 * 1 ) )}
T4_g[ iS14{( i4 * ( 1 * 1 ) )} ]
 root domain : (iS10{i4},bS11{1},bS12{1})
  Merge: bS11{1} and bS12{1} -> bS13{( 1 * 1 )}
  Merge: iS10{i4} and bS13{( 1 * 1 )} -> iS14{( i4 * ( 1 * 1 ) )}
}

In this example bS17 is exact mapped to bS13 but is not permissive mapped to bS13, which was the part that confused the indexing pass.

Semantically, depending on how we want to define permissive mapping, it might make sense to say all of bS13, iS10, and iS14 permissively mapped to bS17, but we might lose precision and consistency quite quickly once we start to map iterdomains on the same tensordomain to each other. This would be the formalization part that'd need some follow up work.

[csarofeen]

WAR LGTM, maybe we should file an issue to track the questions on formalization of permissive mapping.

[naoyam]

LGTM as well