PARQUET-385 PARQUET-379: Fixes strict schema merging

[liancheng]

This PR fixes strict mode schema merging. To merge two PrimitiveType t1 and t2, they must satisfy the following conditions:

1. t1 and t2 have the same primitive type name
2. t1 and t2 either
   • don't have original type, or
   • have the same original type
3. If t1 and t2 are both FIXED_LEN_BYTE_ARRAY, they should have the same length

Also, merged schema now preserves original name if there's any.

[liancheng]

Not quite sure whether we should also deprecate the strict argument of MessageType.union. I tend to do so since Parquet itself never uses non-strict schema merging internally. Also, non-strict schema merging doesn't make sense and may lead to wrong query result as discussed in PARQUET-379.

[lw-lin]

Since originalType is an Enum, seems we can just do like:

if (getOriginalType() != toMerge.getOriginalType())
  reportSchemaMergeError(toMerge);

which covers:

1. if both are null, the condition evals to false
2. if one is null and the other is not null, the condition evals to true
3. if both are not null, then the condition evals to false if they take different values, and evals to true otherwise.

@liancheng what do you think?

[liancheng]

Thanks! Being in the Scala world for so long a time, I really need to re-learn Java again :)

[lw-lin]

Having examined all existing 20 OriginalTypes, and 2 OriginalTypes to be added, I think this PR works well.

So:
+1, LGTM.

[julienledem]

+1

[julienledem]

@liancheng can you remove the square brackets in your pr title? That trips up the merge script.

[liancheng]

@julienledem Done.