docs: add research on eliminating remaining cons

Three constraint-anchored analysis agents explored:

1. Killing T: Pod requirement — VERDICT: Keep Pod. Relaxing to
   Copy+Send+'static causes UB from uninitialized padding bytes on
   the writer side. A NoUninit bound works but adds complexity with
   marginal benefit since DeriveMessage already solves ergonomics.

2. Killing power-of-two capacity — VERDICT: Killable via Lemire
   fastmod (~1.5ns, 2 multiplies). Use is_pow2 branch so power-of-two
   users get zero regression. Const-generic API as opt-in secondary.

3. Killing remaining cons:
   - Async: separate crate, medium effort
   - Pipeline WaitStrategy: easy, 2-4 hours, additive API
   - Derive generics: medium, phased approach (attribute opt-in first)
   - Metrics: separate crate, easy
   - Loom testing: medium, 1-2 weeks

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: userFRM