[bug] CudfNestedLoopJoin drops build byte limit and rejects two-sided timestamp filters

[sperlingxx]

Highlight (by human-in-loop)

1. These are valid tests. Production code regressed; this is not a spark-flux-authored invalid CASE/eq dialect issue.
2. buildByteLimitAndCleanup was added with bounded NLJ accounting in 73a13076e (feat(cudf): Harden MPP exchange and bounded operator execution). The cpp implementation disappeared in the Facebook-main sync #38; the header/config/test still described the limit.
3. innerJoinWithTimestampFilter was added in 635816cba (Improve cuDF coverage for Spark queries). Two-sided gt(p_ts, b_ts) is still valid; AST/JIT gates TIMESTAMP so the compiler tried to precompute the whole predicate and threw sideIdx == -2.

Problem

Part of umbrella #100 (velox_cudf_nested_loop_join_test).

• CudfNestedLoopJoinTest.buildByteLimitAndCleanup: expected an exception (CudfNestedLoopJoin build exceeds configured device-memory limit), but none was thrown (NestedLoopJoinTest.cpp:100).
• CudfNestedLoopJoinTest.innerJoinWithTimestampFilter: VeloxRuntimeError INVALID_STATE: Expression spans both join sides and cannot be precomputed: gt(ROW["p_ts"],ROW["b_ts"]); sideIdx was -2 (AstExpressionUtils.h:523).

Root cause

Byte limit. CudfNestedLoopJoin.h still declares reserveBuildBytes / releaseBuildBytes / maxBuildBytes_, and cudf.nested_loop_join.max_build_bytes still exists. CudfNestedLoopJoin.cpp stopped calling them after #38 replaced the cpp with the upstream operator. The join retained the full build with no cap.

Separately, Values stays on CPU. Default CudfFromVelox coalescing (gpu_batch_size_rows=100000) merges the two tiny build pages into one GPU batch, so even a restored cap would not exercise “first batch retained, second trips the limit, close releases accounting.”

Timestamp filter. containsAstUnsupportedType keeps TIMESTAMP off the AST/JIT path. gt(p_ts, b_ts) then goes through compileSubExpression(), which cannot precompute an expression that spans both join sides. cuDF AST comparison operators still accept timestamp column_references, so the join filter can be compiled from native field refs.

Expected behavior

• With cudf.nested_loop_join.max_build_bytes=16, the second build GPU batch throws the configured device-memory limit, and close releases already-accounted bytes.
• Inner NLJ p_ts > b_ts matches DuckDB.

Acceptance criteria

• Restore build-side byte accounting in CudfNestedLoopJoin.cpp (including close-time release and concat-failure release).
• Compile two-sided timestamp comparisons as AST ops instead of cross-side precompute.
• Pin velox.cudf.gpu_batch_size_rows=1 on buildByteLimitAndCleanup so the two Values pages stay distinct GPU batches.
• Both cases pass, and the rest of velox_cudf_nested_loop_join_test still passes.

Validation

Reproduced and verified on HPDA dev 28798f623 (CUDA SM 86):

• unfixed: no throw on the byte cap; timestamp filter throws sideIdx == -2;
• fixed: both cases PASSED; full binary 58/58.

Prepared with assistance from Cursor.

[sperlingxx]

Fix PR: #112