Write nested lists correctly (#5)

[CGMossa]

TODO: replace this line with your own framing, in your own voice.

AI-written details

Summary

• as_item() chose between an array and a table from whether the key was named rather than from the value list's own names, so any unnamed list passed as a value took the table branch. as_kv_pairs() then inserted every element under the literal key NA, keeping only the last one: as_toml(list(x = list(1, 2, 3))) produced [x] / NA = 3.0. It is now derived from the value's names, mirroring what as_value() already does.
• A list of named lists is written as an array of tables, so an item read with get_item() can be inserted back with insert_items(). This addresses the second half of handling nested list types #5 ("there isn't a clear way to massage this output as something that can be easily inserted back in as an array of tables").
• as_kv_pairs() errors on a non-empty unnamed list instead of emitting NA keys. Unreachable through the public API after the change above, kept as a guard.
• Dropped a set_prefix("\n") decor call on the array branch that was effectively unreachable before and rendered x = followed by a broken multi-line array.
• New tests/testthat/test-nested-lists.R, plus a NEWS entry. Reading these documents already worked after 190531f; the tests cover the read shape too so handling nested list types #5 stays covered end to end.

Test plan

• R CMD check: clean (the only NOTE is a .git entry from building inside a git worktree)
• 20 assertions pass, including the pre-existing formatting snapshots
• The rv.lock v2 snippet from handling nested list types #5 round-trips: get_item() then insert_items() then reparse returns the same data, with the generated-file comments and untouched keys preserved
• The gsm.app example from the issue thread round-trips identical()-equal
• Unchanged: named lists to tables, data.frames to arrays of tables, df_as_array = FALSE, partially-named lists still error

Notes

• Behaviour change: a document holding an inline array of tables (tbls = [{ a = 1 }, { a = 2 }]) is rewritten as [[tbls]] blocks when that value is re-inserted. The data is equal and on main the same operation produced [tbls] / NA = { a = 2 }, so this is strictly better, but it is a formatting change in a package that advertises formatting preservation. Making the array-of-tables branch unconditional was chosen over gating it on df_as_array, whose documented meaning does not extend to lists.
• Two round-trip gaps are left alone, both pre-existing and both read-side representation questions rather than bugs in this path. dependencies = [] reads as NULL, so the key disappears when written back. dependencies = ["lattice"] reads as "lattice" and writes back as a scalar string rather than a one-element array, which is a type change a consumer like rv would reject. Fixing either means changing what get_item() and from_toml() return for length-0 and length-1 arrays for every user. Worth deciding separately, and until then editing a real rv.lock and writing it back is still not safe.

---

Drafted by Claude (claude-opus-5). Reviewed by the author.