feat: align encoder with TOON spec v4.1 (v0.7.0)

[coseto6125]

Upstream moved the spec into its own repo (toon-format/spec) and shipped v4.0 (breaking) and v4.1 since etoon's v3.1 alignment. This PR implements the encoder-side changes and swaps our fixtures for the official suite.

New forms

Nested field groups (§9.3) — a column of uniform objects collapses into the header:

orders[2]{id,customer{name,country},total}:
  1,Ada,DK,99
  2,Bob,UK,149

Keyed tabular (§9.5) — an object of uniform objects becomes a table whose rows carry their own keys:

servers[2:]{host,port}:
  alpha: a.example.com,8080
  beta: b.example.com,9090

Keyless [N:]{…}: at the document root; never for array elements, which stay anonymous (§10).

Encoded-size reduction against the previous nested output: 76.6% (nested field groups) and 31.9% (keyed tabular) on the benchmark payloads.

Fixes

Issue	Spec
Tabular form is mandatory when detection holds, except in list-item position where a keyless fields-bearing header is invalid — that case emitted an illegal - [2]{x}: header	§9.3, §9.4, §6
"+1" was emitted unquoted and decoded back as a number	§7.2
Brackets and braces were only checked at position 0, so x[1] went out unquoted	§7.2

Fixtures

tests/fixtures/encode/ now tracks the official toon-format/spec suite (MIT, 179 encode cases). 178 pass; the single skip needs a non-default indentSize, which etoon hardcodes to 2. key-folding.json stays etoon-local — v4.0 removed key folding from the spec, so fold_keys is an etoon extension now. See ATTRIBUTION.md.

Upstream's rationale for removing key folding is in .out-of-scope/key-folding.md: 0.00% token savings on the reference benchmarks, wire ambiguity against literal dotted keys, and incompatibility with streaming decode.

Verification

• cargo test --no-default-features — 32 pass
• pytest tests/test_spec_fixtures.py — 197 pass
• cargo clippy --features python --all-targets -- -D warnings — clean
• benches/compare.py — output byte-identical to @toon-format/toon 4.1 on all six payloads (toons and py-rtoon now differ on the two v4.1 shapes; they are still on spec v3.x)

Performance (Rust core, A/B against the previous encoder): flat tabular unchanged (155.9 µs → 156.8 µs), deep nested unchanged. Detection costs ~4% on payloads with small object fields that do not qualify (mixed_500: 130.8 µs → 135.9 µs); a first-row column probe rejects those in O(columns) instead of walking every row, which brought that number down from +13%.

Breaking output change, so this is a minor bump to 0.7.0.