Fix: empty array doesn't override an array of tables in toml_merge

[afonsojanu]

`is_array_of_tables()` returns true for an empty array as well, since a loop over zero elements never runs into one that fails the `TOML_TABLE` check. On its own that's harmless, but `datum_merge()`'s `TOML_ARRAY` case uses "both sides are arrays of tables" as the signal to append src onto dst instead of overriding it, and an empty array satisfies that check vacuously.

So merging in a document that explicitly resets an existing array-of-tables key to an empty array doesn't clear it like it should. The merge takes the append branch, appends zero elements from the empty array, and leaves dst's original array of tables untouched:

const char *s1 = "[[arr]]\nx = 1\n";
const char *s2 = "arr = []\n";

toml_result_t r1 = toml_parse(s1, strlen(s1));
toml_result_t r2 = toml_parse(s2, strlen(s2));
toml_result_t merged = toml_merge(&r1, &r2);

toml_datum_t x = toml_get(merged.toptab, "arr");
printf("%d\n", x.u.arr.size);
// prints 1, still [{x=1}], even though r2 asked for arr = []

Excluding the empty-array case from `is_array_of_tables()` makes that combination fall through to the plain override path instead, same as any other src value that isn't a genuine array of tables. I checked this doesn't change behavior for any of the other size/type combinations the previous `is_array_of_tables(*dst) && is_array_of_tables(src)` check (from #44) was meant to distinguish, since an empty array on either side that's genuinely being appended-to or replaced by a real array of tables gets the same end result either way.

Added two cases to `test/merge/test1.c`: one for the array-of-tables-overridden-by-empty-array direction (fails without the fix), and the reverse direction as a sanity check that it still works.

Ran `make clean test` locally; everything passes except `test-stdtest`, which needs the `toml-test` Go binary that wasn't available in my environment and is unrelated to this change.

[cktan]

Thanks! This landed on main as ae0ce2f (squash-merged locally, so GitHub shows the PR as closed rather than merged). Authorship is preserved on the commit.