bench: add large_dump.json (~20 MB) and deep_nest.json fixtures

[membphis]

Background

The bench harness (benches/lua_bench.lua) currently ships with only two fixtures:

• benches/fixtures/small_api.json (~2 KB)
• benches/fixtures/medium_resp.json (~100 KB)

The README roadmap (spec §9.3) calls out two additional fixtures that would expose scaling behavior not visible on the current ladder:

Fixture	Target size	Intent
large_dump.json	~20 MB	Bandwidth-bound behavior; separates L3-cached from memory-bandwidth-limited throughput
deep_nest.json	depth stress	Worst-case for Phase 2 cursor walk and validate_brackets; exposes skip-cache misses

Acceptance criteria

• Both files committed under benches/fixtures/.
• make bench includes them in the size ladder without special flags.
• large_dump.json should be a realistic multimodal or API-response payload (not random bytes), so the bench reflects actual usage patterns.
• deep_nest.json should exercise at least 200 levels of nesting.

Notes

• Synthetic generation scripts (Python / shell) are fine; commit the generated JSON, not the generator, so CI is deterministic.
• deep_nest.json will also stress validate_brackets fusion (see related issue on AVX2/NEON fusion) and the memchr2 jump (very long string interiors). Having the fixture available is a prerequisite for measuring both.
• Check the git lfs policy before committing a 20 MB file; if LFS is not configured, generate it at make bench time from a smaller seed instead.