Canonicalize nested template format specs in annotations - #9076

Summary

Extract the independent template format-spec unparse correction from #8954. Interpolation format specs contain ordinary formatted-value fields and must use the existing f-string body unparser. The outer template interpolation continues to preserve its source spelling.

For example, a future annotation containing t'{value:{(width)}}' is canonically stored as t'{value:{width}}'. This one-line dispatch fix matches CPython 3.14.7 and 3.14.8.

No broader interpolation text, debug-comment, or newline reconstruction changes are included. Checked adjacent #8160 and other open codegen work; this is the separate format-spec dispatch.

Validation

• Independent source review and exact-diff review.
• Existing Python annotation snippet gains three compile-only cases: literal format spec, nested redundant parentheses, and preserved outer spelling with nested conversion.
• Full snippet passes on CPython 3.14.7 and 3.14.8; the prior native binary fails the new assertion. Patched native snippet and both selected pytest cases pass.
• Unchanged canonical test_unparse.UnparseTestCase.test_tstrings and existing structural future-template annotation test pass.
• Normal hooks, native build, codegen/VM and separately configured C-API Clippy with warnings denied.

The regression only compiles and inspects annotation constants; it never executes the generated annotation or formatting callbacks. Local validation is Linux; the full hosted matrix remains required.