perf: single-pass addMeta() stats, Set-based filter() and addRelation…

[bassim-matar]

…s(), indexed-loop createIndex() (0.12.6)

Boot-path and batch-mutation performance, identical behavior:

• addMetaVariables(): one row-major pass with per-column accumulators instead of two column-major scans per column (~2.3x on 100k rows x 29 cols), preserving the empty-string and NaN type-inference edge cases, now covered by tests
• filter(): Set lookups instead of array includes, linear instead of quadratic (~1000x when deleting 10k of 50k rows with cascade)
• createIndex() family: indexed loops instead of Object.entries() + parseInt() (~1.65x on 200k rows)
• planRelations() / appendRelationFieldValues(): Set-tracked ids, linear batches (~75x for a 10k-id addRelations call)
• getAllChilds(): push accumulation instead of per-level concat
• minor: loop instead of spread-reduce in loadTables(), Set in checkConformity()