feat(harness): support per-space skill policies

[Rhosmarie]

Summary
Support per-Skill-Space allow/deny policies in Harness selected skills.

This lets callers load a Skill Space without downloading every skill in it, for example:

selected_skills=[
    HarnessSelectedSkill(
        source="skillspace",
        skill_space_id="ss-a",
        skill_space_policy=HarnessSkillSpacePolicy(mode="allow", ids=["s-a1"]),
    ),
    HarnessSelectedSkill(
        source="skillspace",
        skill_space_id="ss-b",
        skill_space_policy=HarnessSkillSpacePolicy(mode="deny", ids=["s-b9"]),
    ),
]

Changes

• Add HarnessSkillSpacePolicy and HarnessSelectedSkill.skill_space_policy.
• Support per-space SKILL_SPACE_POLICY format while keeping the existing global {mode, ids} format compatible.
• Apply policy filtering before materializing/downloading remote Skill Space skills.
• Pass request-level Harness policy explicitly instead of temporarily mutating os.environ, avoiding concurrent request policy leakage.
• Treat mode="allow", ids=[] as a valid “disable all remote skills in this space” policy.
• Reject conflicting policies for the same skill_space_id.
• Add regression coverage for allow/deny filtering, empty allowlist, env compatibility, Harness spawn, and sandbox forwarding.