fix: preserve mandatory read and post-join filters

[bvolpato]

Mandatory ReadRel filters and JoinRel post-join filters are currently dropped when converting plans to Calcite or Spark. A read carrying FALSE therefore returns its input rows, and an outer join loses predicates that should inspect its null-extended output.

For example, this scan must produce no rows, but Isthmus currently returns an unrestricted TableScan:

NamedScan scan = NamedScan.builder()
    .addNames("t")
    .initialSchema(NamedStruct.of(
        List.of("id"), TypeCreator.REQUIRED.struct(TypeCreator.REQUIRED.I32)))
    .filter(ExpressionCreator.bool(false, false))
    .build();
new SubstraitToCalcite(ConverterProvider.DEFAULT).convert(scan);

Apply mandatory scan filters before emit remapping and post-join filters after the join, using the actual output schema. Decode protobuf post-join references against that same schema so outer-join nullability is preserved and IS NULL cannot be incorrectly simplified away. These semantics follow spec v0.102.0.

Partially addresses #1204 for mandatory read filtering.

[alexandrefimov]

Heads-up on an overlap: ReadRel.projection, one of the three fields #1204 names, is in #1280 -- it applies the mask in visit(NamedScan) and visit(VirtualTableScan), where your applyFilter calls also sit.

The two compose: the filter stays on the scan, the mask goes above it, the emit mapping above that -- a filter is read against the direct schema, before the projection (spec v0.102.0, Read Filtering). Whichever of us lands second rebases, and I am glad to be that one.