feat: Implement workflow execute() methods in terraphim_multi_agent

Problem

MultiAgentWorkflow::execute() in terraphim_multi_agent/src/workflows.rs is a placeholder that returns todo!() for all 5 workflow patterns:

• RoleChaining: sequential agent execution, output of N feeds input of N+1
• RoleRouting: condition-based routing to specialist agents
• RoleParallelization: concurrent agent execution with result aggregation
• LeadWithSpecialists: lead agent delegates subtasks, aggregates results
• RoleWithReview: proposal agent + review agent with feedback loop

The enum and pattern definitions exist but none execute. This blocks any actual multi-agent workflow orchestration.

Proposed Change

Implement actual workflow execution for each pattern:

1. RoleChaining: Chain agents sequentially. Each agent receives the previous agent's output as context. Final output is the last agent's response.
2. RoleRouting: Evaluate routing conditions against input, dispatch to the matched specialist agent.
3. RoleParallelization: Spawn all agents concurrently via tokio::join_all, aggregate results (configurable: merge, vote, or pick-best).
4. LeadWithSpecialists: Lead agent decomposes task, delegates subtasks to specialists, receives results, and synthesizes final output.
5. RoleWithReview: Proposal agent generates output, review agent critiques, loop until approval or max iterations.

Scope

• crates/terraphim_multi_agent/src/workflows.rs (currently 84 lines, expected expansion to ~400-500 lines)
• Unit tests for each workflow pattern

Context

This is UPLIFT-6 from the medgemma-competition multi-agent integration plan. These workflow patterns are essential for clinical reasoning pipelines (e.g., RoleChaining for diagnosis->PGx->treatment, RoleWithReview for safety validation).

Related upstream issues: #534, #535, #536

Start by reading terraphim_multi_agent/src/workflows.rs and the existing workflow pattern and enum definitions. Implement and test execute() for RoleChaining, RoleRouting, RoleParallelization, LeadWithSpecialists, and RoleWithReview; done means each pattern runs its described orchestration and has a unit test.