Add registry mappings and CI profiles for AI/ML catalog

[ulises-jeremias]

Problem

Adding AI/ML templates and extensions will break CPA CI unless the registry validation
scripts and curated profiles know the new template types and compatibility rules.

CPA's CI is layered: L0 validates registry/profiles, L1 scaffolds templates alone, L2
tests one extension on its canonical template, and L3 tests curated profiles. New
AI/ML types must fit that flow.

Solution

Create a registry and CI support issue for the AI/ML catalog.

Deliverables:

• Add new categories to templates.json only after taxonomy is accepted.
• Add new template type mappings in:
  • scripts/ci/registry.py::CANONICAL_TEMPLATE_BY_TYPE
  • scripts/ci/validate-registry.py::STACK_PREFIX_BY_TYPE
• Add CI profiles under ci/profiles/ for realistic stacks, for example:
  • fastapi-ai-chat-default
  • mlops-sklearn-default
  • mlops-sklearn-github-actions
• Keep L3 profiles small and representative; do not test every compatible extension
  combination.
• Ensure new categories do not cause duplicate-category conflicts in curated profiles.
• Add or update tests for registry validation if current scripts support it.

Constraints:

• No generated project should require network/API credentials to pass CI.
• Heavy GPU/distributed paths must be out of default CI unless explicitly mocked or
  reduced to CPU fallback.
• Every new template/extension issue must specify its expected CI profile.

Acceptance criteria:

• L0/L1/L2/L3 can validate the first AI/ML vertical slice.
• Future AI/ML issues know exactly which registry/CI files must change.

[ulises-jeremias]

CI trust requirements (from #92 / CNA #189–193)

When wiring AI/ML profiles:

1. Every new AI/ML template type gets a bare L1 job even with zero extensions.
2. L2 must run pytest (or the template test command) in the generated project.
3. Do not stack all AI/ML extensions in one job — small curated L3 profiles only.
4. No network/API credentials; use .env.example / fixtures.

See also #91 for incompatibleWith matrix.

[ulises-jeremias]

Reopened after fully reverting #93 via #94. Work remains for the assigned contributor — please continue from this issue.