refactor(config): split step definitions into a discriminated union

[hertznsk]

Closes #517.

What

Replaces the monolithic AgentDef — one model carrying every step type's fields, with a hand-maintained validator matrix deciding which fields each type: accepts — with concrete Pydantic models united by a static StepDef discriminated union:

• AgentDef is now the provider-backed LLM step only.
• New public models: HumanGateStepDef, QuestionsStepDef, ScriptStepDef, MCPStepDef, WaitStepDef, SetStepDef, TerminateStepDef (deliberately no routes), WorkflowStepDef, plus StepBase / RoutableStepBase. All exported from conductor.config and conductor.config.schema.
• WorkflowConfig.agents: list[StepDef] and ForEachDef.agent: StepDef parse into concrete variants through Field(discriminator="type") with an outer BeforeValidator that canonicalizes the legacy shorthand (missing or explicit-null type) to "agent" at the boundary.
• Every variant owns exactly the fields meaningful for its kind with extra="forbid", so a misplaced field is rejected next to its step instead of being silently ignored — this is what makes new step types cheap to add (no validator matrix to extend).
• The published JSON Schema now exposes oneOf + discriminator with per-variant additionalProperties: false, so editors and external tooling get per-type completion and validation.

Compatibility

• Workflow YAML is unchanged. Omitted type and type: null still load as LLM agents (canonicalized to type: "agent" at parse time).
• AgentDef(...) keeps working for LLM agents, including AgentDef(type=None, ...), model_dump(exclude_none=True) round-trips, model_copy(update=...), and the synthetic agents built by OutputValidator and the ACA runner.
• Intentional tightenings: constructing a non-LLM step via AgentDef(type="script", ...) no longer works (use the named step class); a field belonging to a different step type now fails with Pydantic's standard extra_forbidden rather than the custom "<type> agents cannot have '<field>'" messages (variant-owned invariant messages are unchanged); HumanGateStepDef lost the inert model field; human_gate is rejected as a for_each inline agent (concurrent iterations would compete for one interactive channel — route to a gate from the group's routes: instead).

Engine and consumers

• Main / parallel / for-each dispatch narrows via isinstance, with defensive ExecutionError guards for variants the static validator would reject — a directly-constructed engine never runs it. The for-each guard fires before the empty-source early return so a forbidden inline variant can never pass silently.
• Generic consumers narrowed the same way: validator routing graph and subworkflow/template scans, conductor show/run/plugin, the MCP conductor_plan_tree (a terminate step now renders with empty routes), executor skills/plugins helpers, and the synthetic provider agents (max_depth removed).
• FileString (!file) preservation restored on human_gate/questions prompts so file-backed gate prompts keep resolving relative {% include %} paths.

Testing

• Full suite: 9113 passed, 3 failed — all three failures reproduce identically on main in this environment (two chmod(0o000) permission tests and one timing-sensitive performance test); none touch this change.
• ruff check clean; ty check src shows only the same pre-existing diagnostics as main; make validate-examples green.
• New regression coverage: JSON Schema oneOf/discriminator/additionalProperties shape, ForEachDef.agent variant parsing, legacy missing/null-type canonicalization, model_dump round-trips for all nine variants, resume round-trip from a legacy no-type YAML, runtime defensive guards for parallel/for-each groups (including ahead of the empty-source early return), conductor_plan_tree with a terminate step, and an explicit-empty output_template: {} coverage-override regression.