Fix maps:from_keys/2 dedup for boxed terms (tuples, etc.)

[CoderDennis]

Claude and I hit this bug when building what I thought would be a simple Conway's Game of Life on an ESP32 device with a small touch screen.

Expected behavior

maps:from_keys(List, Value) and sets:from_list(List) should deduplicate List using
structural equality — the same semantics as OTP, and as AtomVM's own term_find_map_pos. Five
separately-allocated {1, 1} tuples should collapse to size 1.

Actual behavior

On AtomVM, five separately-allocated but structurally-identical {1, 1} tuples produce a
map/set of size 5 — no deduplication happens for boxed terms. Immediates (small integers,
atoms) are unaffected.

sort_keys_uniq's dedup phase compared keys with raw pointer != after sorting them with real structural comparison. For immediates (small integers, atoms), the tagged word is the value, so != happened to work. For boxed terms like tuples, the word is a tagged pointer, so two separately allocated but structurally equal keys were never deduplicated, even though the preceding sort placed them adjacent via correct structural comparison.

sets:from_list/1 hits the same bug since it's built on top of maps:from_keys/2.

Fix

Use term_compare/4 in the dedup phase instead, mirroring the existing pattern in term_find_map_pos.

Add a regression test to maps_nifs.erl: build five separately heap-allocated {1, 1} tuples and confirm maps:from_keys/2 collapses them to a map of size 1.