Implement the 0.37 performance tranche

[rbroderi]

Summary

• make dense next traversal and deletion sweeps linear with successor-preserving iteration indexes
• broaden allocation-free scalar Python entry to strict float and two-integer leaves
• remove unreachable direct-leaf cell work and preserve generated deletion semantics
• add the reusable 0.37 A/B and three-way LuaPyre/native Lua 5.5/Python lower-bound benchmarks
• bump the package to 0.37.0a1 and document accepted, rejected, and deferred work

Performance

The accepted targets improve by 30.8–97.0% across CPython 3.13 and 3.14. The new headroom corpus records all nine algorithms against native Lua 5.5 (lupa.lua55) and reduced-contract Python with three isolated, Latin-square-ordered processes per version.

Validation

• 592 tests pass on CPython 3.13.15 and 3.14.7
• all 24 required official Lua 5.5.1 probes pass on both versions
• Penlight 23/23, luatest 5/5, LuaCov scanner 24/24, and Are We Fast Yet 11/11 pass on both versions
• benchmark outputs and adaptive code-generation audits are committed