fix(ai): handle custom LLM models that emit test-case declarations

[DavidHoenisch]

When using an override Ollama model (e.g. llama3.2:latest) the LLM often outputs the generated test cases as a Go variable declaration instead of a bare array literal or a complete test function:

tests := []struct {
    name string
    args args
    want int
}{
    {name: "case 1", args: args{...}, want: 42},
}

Two parser bugs caused these responses to be rejected:

1. parseGoTestCases() only recognised bare composite literals or complete func Test... blocks. The tests := form fell through to parseTestCaseArray, which wrapped it in another literal producing invalid Go.

2. ensureTrailingComma() saw the declaration ended with '}' and was not a func Test, so it appended a comma: "tests := []struct{}{...}," which is invalid Go syntax.

Both issues caused GenerateTestCases() to return an error. The error output (which contained the LLM's generated cases) was printed to stderr/stdout, while the actual test file only received the TODO fallback comment.

Fix:

• Detect tests := / tests= declarations in parseGoTestCases and wrap them in a func init() so the existing AST parser can extract cases.
• Teach ensureTrailingComma to skip tests := and var tests declarations.

Verified with llama3.2:latest and qwen2.5-coder:0.5b.