ci: move llama model demo out of PR workflow and fix print expectation

[cpunion]

Summary

• remove the HuggingFace llama model download/artifact dependency from the PR LLGo workflow
• skip _demo/c/llama2-c during the normal demo sweep because it requires an external model checkpoint
• add a separate Model Demo workflow that runs the llama demo once daily and can also be triggered manually
• fix the Go <1.26 builtin print regression tests to expect legacy three-digit exponent padding

Verification

• bash -n .github/workflows/test_demo.sh
• git diff --check
• go run github.com/rhysd/actionlint/cmd/actionlint@latest .github/workflows/llgo.yml .github/workflows/model-demo.yml
• GOTOOLCHAIN=local go test ./test -run TestBuiltinPrintLegacyFloatFormat -count=1
• Docker CI-like Go 1.24.2 linux/amd64:
  • llgo test -timeout=20m github.com/goplus/llgo/test -run TestBuiltinPrintLegacyFloatFormat
  • llgo test -timeout=20m github.com/goplus/llgo/test/go -run TestBuiltinPrintLegacyExponentWidth

This prevents PR CI from failing on HuggingFace 429s while preserving scheduled coverage for the model-backed demo, and fixes the current main CI print-format failure.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!