fix(isthmus)!: preserve literal nullability converting to Calcite

[alexandrefimov]

Substrait literal nullability did not survive conversion to Calcite: RexBuilder gives every non-null literal a NOT NULL type, so nullable literals came back non-nullable, and a homogeneous NestedList holding a nullable literal next to a nullable field reference came back mixed-typed and was rejected. Only the character/binary paths kept the type, incidentally, by picking up a cast wrapper.

This makes that shape the contract. Scalar literal visitors request their exact Substrait type (makeLiteral(value, type, allowCast = true)); the makeTimeLiteral / makeTimestampLiteral / makeIntervalLiteral paths wrap through preserveNullability. Interval-day no longer needs it: #1120 moved that visit onto the type converter, which already carries nullability. On the way back, CallConverters.CAST folds a cast over a non-null literal that changes nothing but nullability into the literal itself, carrying the cast's nullability, instead of surfacing an Expression.Cast. LogicalValues tuples keep bare literals — the row type already declares field nullability there, so that one cast is unwrapped, and only that one: a cast doing anything else is doing work, and dropping it would leave the literal under a row type that does not describe it.

Found along the way and left as is, tracked separately: the Calcite→Substrait path normalizes time/timestamp literals to microsecond precision (#1114) and a precision_timestamp_tz literal returns as a plain precision_timestamp (#1113).

Closes #1066

BREAKING CHANGE: Substrait→Calcite conversion now represents a nullable literal as a cast of the literal to its nullable type where it previously produced a bare literal with a non-nullable type, and converted row types become nullable accordingly. Calcite→Substrait conversion folds nullability-only casts over literals into nullable literals instead of Expression.Cast.

[nielspardon]

With #1120 merged you can update this PR and we can check what needs fixing.

[alexandrefimov]

Rebased. The interval-day half went away rather than merged: #1120 moved visit(IntervalDayLiteral) onto makeLiteral(value, typeConverter.toCalcite(...), true), so the type already carries the nullability there and the preserveNullability wrapper had nothing left to do. What remains is makeTimeLiteral, makeTimestampLiteral and the year-interval makeIntervalLiteral.

The round-trip table keeps its interval_day row. It passes because of #1120 now rather than because of this PR, but it seemed worth keeping as coverage rather than deleting.

Also checked this against the neighbours: #1110, #1121 and #1127 all touch the same visits, and a tree with all of them merged onto main builds. Not something CI can tell us, since it only ever sees them one at a time.

[nielspardon]

Rebase reasoning checks out — visit(IntervalDayLiteral) on main does go through the type converter now, so that half really is subsumed rather than merged. Design verified against Calcite 1.42: makeAbstractCast does no literal folding, so the nullable shape is stable and matches what the scalar visitors produce. Three small things inline.

Separately, the precision_timestamp_tz degradation and the microsecond normalization are recorded only in a test comment — worth issues so they don't get lost.

[alexandrefimov]

Rebased. The core commit is gone: #1136 landed the same one-line fix, and its LiteralTypeDerivationTest covers what the test in that commit covered plus interval_day and interval_compound, so keeping mine would only have duplicated it. The description and the breaking footer no longer mention it.

On the two things recorded only in a test comment — both are filed already, and both are closed by #1121: #1114 for the microsecond normalization, whose interval third went with #1120, and #1113 for the precision_timestamp_tz degradation. The description points at them now instead of leaving them in a comment.

[nielspardon]

LGTM