feat: add sub-agent orchestration with zero-trust permission model

[bug-ops]

Summary

• Sub-agent orchestration system enabling the main agent to delegate tasks to specialized child agents
• Zero-trust security model: all permissions TTL-bounded, auto-revoked, secrets never persisted
• In-process A2A channels (mpsc) reusing zeph-a2a types, upgradable to HTTP later
• Tool filtering (AllowList/DenyList/InheritAll) and skill filtering with glob patterns
• 55+ unit tests covering parsing, grants, filtering, manager lifecycle

New module: zeph-core::subagent

File	Purpose
def.rs	SubAgentDef, TOML frontmatter parser, definition loader
manager.rs	SubAgentManager (spawn/cancel/collect/approve_secret)
channel.rs	In-process A2A channels (OrchestratorHalf/AgentHalf)
grants.rs	PermissionGrants with TTL, audit trail, Drop revocation
filter.rs	FilteredToolExecutor, SkillFilter with glob matching
error.rs	SubAgentError

Zero-trust security model

• Sub-agents start with zero permissions
• Secrets delivered via in-memory grants only (never serialized)
• All grants have TTL with automatic expiry sweep
• Drop impls revoke all grants on sub-agent termination
• Audit trail via tracing for grant/revoke/expire events
• Fenced-block tool calls blocked for sub-agents (SEC-03 fix)

Test plan

• cargo +nightly fmt --check
• cargo clippy --workspace -- -D warnings
• cargo nextest run --workspace --lib --bins (2294 passed)
• Security audit: 0 critical, 0 high remaining
• Code review: APPROVED

Closes #709, #710, #711, #712, #713