fix: suppress misleading "Did you mean" suggestions

[anmarhindi]

Summary

suggestCommand previously printed a suggestion for every unknown subcommand, even when the best jaro-winkler match was barely related. Two visible failure modes:

$ openai responses zzzzz
No help topic for 'zzzzz'. Did you mean ''?

$ openai totallybogus
No help topic for 'totallybogus'. Did you mean 'openai completions'?

Add a minimum similarity threshold below which suggestCommand returns "". The upstream urfave/cli error formatter already omits the suggestion clause when the function returns an empty string, so the user just sees No help topic for 'X'. cleanly.

After

$ openai responses zzzzz
No help topic for 'zzzzz'

$ openai totallybogus
No help topic for 'totallybogus'

$ openai responses creat
No help topic for 'creat'. Did you mean 'openai responses create'?

$ openai chat:completion
No help topic for 'chat:completion'. Did you mean 'openai chat:completions'?

Why 0.7

I measured jaro-winkler scores against the real top-level commands. Real typos score 0.78-1.00 (e.g. creat -> create = 0.97, comp -> completions = 0.87, lst -> list = 0.93). Unrelated input scores at most 0.557 (totallybogus -> completions). A threshold of 0.7 sits in the gap, and conveniently matches the existing boostThreshold constant inside jaroWinkler itself — the prefix boost only kicks in past 0.7, so it's a natural "plausibly the same word" cutoff.

Verification

• go test ./pkg/cmd/ -run TestSuggest — new table-driven coverage of close typos, exact matches, unrelated input, low-similarity input, empty input, and empty command lists.
• Manually exercised the four cases above against a freshly built binary.

Closes #16