perf(grid): compile filter predicates once instead of per row (#1550)

[ZhuchkaTriplesix]

Summary

_PredicateNode.evaluate re-derived everything from the raw strings for every row: op.toUpperCase().trim(), targetValue.toLowerCase() and double.tryParse(targetValue) (the target never changes), the lower-cased and trimmed IN list on each row (O(rows x list)), and the BETWEEN bounds.

The constructor now resolves them once: _upperOp, _lowerTarget, _numTarget, _inSet (a Set<String> of normalised values) and _numMin / _numMax. evaluate is a switch on the resolved operator and only reads the cell. Semantics are as before: numeric comparison when both sides parse, otherwise case-insensitive string comparison; NULL / null / empty count as null; BETWEEN without both bounds falls through to the comparison path.

The only per-row string work left is on the cell side (toLowerCase, trim, tryParse of the cell), which cannot be hoisted.

This touches the node class only; the lexer changes in #1607 are in the same file but different lines.

Tests (not run locally; CI will)

test/features/workspace/grid_filter_predicate_test.dart: IN / NOT IN with mixed case and spaces, numeric and string BETWEEN, numeric vs string comparison (20.0 = 20, price = abc), : and != ignoring case, and IS NULL for NULL / null / empty.

The before / after timing on 50k rows asked for in the issue is not included; the harness is not in the repo.

Closes #1550