perf(sort,filter): Schwartzian transform and remove recursive match

[rkaraivanov]

sort:

• Pre-compute sort keys once per item (O(n)) instead of on every comparison (O(n log n)) via Schwartzian transform
• Store decorated rows as flat pre-allocated tuples [item, key0, key1, ...] to halve allocations and improve cache locality vs {item, keys:[]}
• Pre-compute direction multipliers array to eliminate map lookups from the comparator hot path

filter:

• Replace recursive match + per-record ands.filter() with matchTree, eliminating O(n x ors) array allocations
• Pre-compute ors/ands arrays per expression tree once per apply() call instead of re-deriving them per record

Profiling the demo grid with templates and 100k records:

• sort ~2x faster than the previous implementation
• filtering ~3-5% faster but with less memory allocations/GC pressure during the operation