feat(harness): surface skill-load errors + per-invoke max_llm_calls

[yaozheng-fang]

What

Two harness-server fixes/additions.

1. Skill-load errors are surfaced (not silently skipped)

Previously build_skill_toolset caught any per-skill failure and only logged a
warning, leaving the agent with no skills while ADK's SkillToolset still
told the model it had them — so the model answered "I can't find that skill."
(Most commonly the skill failed ADK's Frontmatter validation, e.g. its
description exceeds the hardcoded 1024-char limit.)

Now it fast-fails with a SkillLoadError naming the skill and the reason:

• a base skill (from harness.yaml) that fails aborts server startup, so
  a bad config surfaces at deploy instead of shipping a skill-less runtime;
• a once-time override skill that fails is caught in /harness/invoke and its
  reason is returned to the caller (HTTP 200, in the response output).

2. max_llm_calls, threaded into the runner's RunConfig

• A harness default: harness.yaml → MAX_LLM_CALLS env → HarnessConfig.
• A per-invocation override: run_agent_request.max_llm_calls.

Both flow into runner.run(run_config=RunConfig(max_llm_calls=...)). CLI:
veadk harness add --max-llm-calls N and veadk harness invoke --max-llm-calls N.

Verified locally

• Override skill clawhub/heardlyapp/first-class-usps (desc > 1024) → invoke
  returns Skill '...' failed to load: ... description must be at most 1024 characters.
• Valid skill clawhub/harrylabsj/route-weaver → loads; the model calls
  load_skill and follows the skill workflow.
• A bad base skill → server startup raises SkillLoadError.
• max_llm_calls: per-call 2 / 7 and the harness default 50 all appear in
  the runner's run config; the per-invoke value takes effect dynamically.
• ruff + pyright clean.

🤖 Generated with Claude Code