fix(strategies): describe(id, registry) FieldError for 3+ type-parameter strategies

[ocots]

Summary

• _strategy_base_name and _strategy_type_name each peeled exactly one UnionAll layer before assuming a DataType underneath, which only holds for strategies with exactly two total type parameters. CTSolvers.Integrators.SciML (4 params) crosses that line, breaking describe(:sciml, ...) and, because the parameter branch walks every strategy in the registry, describe(:cpu, ...) / describe(:gpu, ...) too on any registry containing it.
• Both helpers now handle any parameter depth: _strategy_base_name falls through to the existing nameof-based fallback, _strategy_type_name peels every UnionAll layer instead of one.
• Adds a new arity-matrix test (1–4 type parameters) exercising the public describe(id, registry) entry point, plus a describe(:cpu, registry) regression covering the blast-radius case — this entry point had zero prior coverage.
• Adds a describe(:di, registry) test for DifferentiationInterface, CTBase's own strategy, previously only reachable through CTFlows' flow registry.
• Adds validation in create_registry for a related, previously-silent contract: the strategy parameter must be the first declared type parameter (every Strategies.parameter implementation reads slot 1). Violating it used to surface as a bare TypeError or a misleading NotImplemented; it now raises a clear Exceptions.IncorrectArgument.
• Documents the parameter-position contract in AbstractStrategy's docstring.

Fixes #516

Test plan

• suite/strategies (980 tests) green
• suite/differentiation (108 tests) green
• suite/meta code-quality/docstring checks green
• Full suite green: 4923/4923

🤖 Generated with Claude Code