ci: make integration tests more frugal

[davidberenstein1957]

Problem

Each authorized PR runs ~21 integration jobs that hit live LLM APIs. Recent runs showed:

• Redundant provider coverage (google + gemini, openai + bare, multiple Azure aliases)
• Triple Python version matrix (3.12–3.14) despite Main CI already running unit tests on all three
• A both agents backend job that re-runs the union of giskard-llm and litellm suites
• A checks job that runs make test-functional but collects zero tests (giskard-checks uses @pytest.mark.integration, not functional)

Failures are often infra-related (e.g. OpenAI credits exhausted) rather than regressions, while API cost scales with every PR.

Proposal (phase 1 — workflow only)

Shrink the integration matrix without changing test code:

1. Python 3.12 only for integration jobs (unit tests keep 3.12–3.14 on Main CI)
2. Drop alias LLM providers — keep openai, google, anthropic, azure; drop bare, gemini, azure_ai, azure_foundry_v1
3. Drop both agents backend — giskard-llm and litellm jobs already cover both install layouts
4. Remove dead checks integration job until checks uses functional marks or a dedicated smoke tier

Expected impact: ~21 jobs → 6 jobs, ~70% fewer live API calls per PR.

Follow-ups (phase 2+)

• Introduce @pytest.mark.functional_smoke for PR CI vs full suite on nightly/main
• Path-filter integration workflow to LLM/agents changes only
• Wire giskard-checks integration tests (--run-integration) into a smoke job when ready

[linear-code]

OSS-111

[davidberenstein1957]

Phase 3 (path filtering) is now included in #2696:

• Workflow-level paths — skips integration entirely on docs-only / unrelated PRs
• detect-changes job — agents-only PRs skip the 7 llm provider jobs; llm/core changes still run both suites
• workflow_dispatch — always runs the full matrix for manual verification