Reduce per-row sorting allocations in multi-entry map_sort normalization #5900

Normalizing multi-entry map<string,int> batches currently slices each row's keys and invokes Arrow sorting, allocating temporary indices and prefix tuples for every map. The singleton optimization in #5887 avoids this work for one-entry maps, but multi-entry maps still pay these costs.

Reuse an Arrow-compatible prefix-tuple scratch buffer across rows for multi-entry string-key/int-value maps, while preserving the singleton fast path, exact duplicate-key permutations, sliced offsets, null buffers, schema metadata, and unsupported-key errors. Also fill rebased offsets in bulk for all-empty visible batches.

Matched release microbenchmarks show about 3x faster normalization for 2–10-entry forward maps and 33–39% faster normalization for 2–50-entry maps in independent paired measurements. Validation covers normalization, hashing, combined execution, nulls, mixed cardinalities, Unicode keys, and fallback types.

Related to #5818 and the matched benchmark work in #5822.