fix(schema): decode Python string literals in static schema

[breken-ai]

Summary

• Make parseStringLiteral (static schema generator) return the value Python gives a string literal: strip r/u prefixes and quotes, decode backslash escapes in non-raw strings, and join implicitly concatenated literals. Bytes and f-strings are still rejected.
• Add table tests for the literal parser and a parser test that checks description, regex and default of an Input(...).

Why

parseStringLiteral returned the source text between the quotes. The bundled .cog/openapi_schema.json then differs from the predictor's real values:

code: str = Input(
    description="A numeric code. "
    "Digits only.",
    regex="^\\d+$",
)

On main this gives:

• description: A numeric code. "\n "Digits only. The inner quotes and indentation end up in the schema and API docs. Splitting long descriptions across lines like this is common.
• pattern: ^\\d+$ (two backslashes) instead of ^\d+$. coglet validates inputs against this schema, so "123" is rejected even though the predictor's own regex accepts it.
• r"""...""" came back as "".."", and u"..." was not recognised. For default=, that made the default None.

Test plan

• go test ./pkg/schema/python -run 'TestParseStringLiteral|TestInputStringArgumentsUsePythonStringValues' -count=1 fails on main and passes with this change
• go test ./pkg/schema/... ./pkg/config/... -count=1 (includes the fuzz seed corpora)
• go build ./...
• golangci-lint run ./pkg/schema/... (v2.10.1): 0 issues

I found and fixed this with an AI coding agent, and I checked the change and the test results above.