perf: rewrite stage_1 lexer with custom(), optimize stages 2-4

[NTBBloodbath]

Replace chumsky combinator-based lexer with hand-written loop inside chumsky's custom() escape hatch. Coalesce consecutive Regular chars into Text(String) runs to reduce downstream token count. Eliminate wasted Vec allocations in stage_2 whitespace separators. Remove per-item clone() in stage_3 dedup_opener_candidates; fuse post- rollup unravel+eliminate into single pass. Convert stage_4 to slice-based traversal, eliminating recursive to_vec() allocations.

Stacked on top of #30

---

Benchmarks vs pre-migration baseline (chumsky 0.9.3, all LTO):

Fixture: small (747B)
parse_tree: 2.54ms → 0.25ms (10x)
parse_flat: 2.57ms → 0.23ms (11x)
stage_1: 1.93ms → 12µs (161x)
stage_2: 71µs → 52µs (1.4x)
stage_3: 584µs → 162µs (3.6x)
stage_4: 24µs → 12µs (2.0x)

Fixture: medium (50.8KB, 1108 lines)
parse_tree: 151.7ms → 10.6ms (14.4x)
parse_flat: 149.5ms → 10.2ms (14.8x)
stage_1: 128.5ms → 0.89ms (144x)
stage_2: 2.33ms → 1.46ms (1.6x)
stage_3: 22.1ms → 7.69ms (2.9x)
stage_4: 1.33ms → 0.59ms (2.3x)

Fixture: large (6KB, 204 lines)
parse_tree: 18.5ms → 1.68ms (11x)
parse_flat: 18.0ms → 1.59ms (11x)
stage_1: 14.0ms → 100µs (140x)
stage_2: 422µs → 282µs (1.5x)
stage_3: 3.34ms → 1.17ms (2.9x)
stage_4: 256µs → 101µs (2.5x)

All 31 tests pass. Zero warnings (build + clippy).

Key changes:

• src/stage_1.rs: custom() lexer replaces chumsky combinators. Adds NorgToken::Text(String) variant. Coalesces Regular char runs into Text tokens reducing stage_2 input 5-10x.
• src/stage_2.rs: Replace wasted whitespace .collect::<Vec<_>>() with .ignored() in separators. Add Text(String) arm to tokens_to_paragraph_segment.
• src/stage_3.rs: Replace coalesce() in dedup_opener_candidates with manual fold avoiding per-item 160-byte clones. Fuse post-rollup unravel+eliminate_invalid_candidates into single pass. Remove unused eliminate_invalid_candidates function.
• src/stage_4.rs: Introduce stage_4_from(&[NorgASTFlat]) taking slices. consume_heading_content and consume_nestable_* return (content, new_position) tuples instead of mutating &mut usize. Eliminates 6 recursive flat[i..j].to_vec() allocations.