misc: Fix Projector stream offset mapping for multi-column projection

[xiaoxmeng]

Summary:
CONTEXT: When projecting multiple top-level columns (e.g., int_traits_map and
long_traits_map), the Projector crashes during deserialization because the
projected schema's stream offsets don't match the actual data layout.

The root cause: buildProjectedSchema assigned output offsets sequentially via
depth-first traversal (all int_traits_map streams first, then long_traits_map).
But project() copies input streams to output positions based on
inputStreamIndices_ (sorted by input offset). When sibling subtrees' input
offsets interleave numerically (e.g., long_traits_map nulls=2 falls between
int_traits_map nulls=1 and its first child at offset 3), the depth-first
traversal assigns wrong output offsets — causing the Deserializer to
misinterpret stream data types and crash.

WHAT: Replace sequential offset allocation in buildProjectedSchema with an
offset map (input offset → output index) derived from sorted
inputStreamIndices_. This ensures projected schema offsets exactly match the
data layout produced by project(), regardless of how input stream offsets
interleave across sibling subtrees.

Differential Revision: D95328184

[meta-codesync]

@xiaoxmeng has exported this pull request. If you are a Meta employee, you can view the originating Diff in D95328184.