Fix interpolated string warning diagnostic locations - #9314

Summary

Rebased #9314 onto main fd3c3fca644b93d0bec719efb3f864056b5d3b68 and reviewed its implementation against CPython 3.15.0 source d220409a13b77de7a6c30abe16ea90db01528edc.

• Reconstruct CPython decoder fragments, including named Unicode boundaries and the different doubled-brace rules for body versus format-spec constants.
• Keep source offsets for deduplication and Single-input boundaries, with separate END/MIDDLE token anchors for diagnostic locations and parser action ordering.
• Match the Unicode decoder's protection of non-ASCII characters after backslashes; independently collect tokenizer-owned brace warnings.
• Preserve parser/tokenizer phase order across nested f/t strings, format specs, numeric events, and warning escalation/recovery.
• Preserve pending Single and rejected-type-comment errors while allowing the appropriate tokenizer recovery events. TYPE_COMMENTS uses the shared warning collector instead of prematurely emitting format-spec callbacks.
• Match numeric zero-width error metadata and physical-line text, and ordinary Unicode-token final-line text and universal-newline positions, including exec bare CR.

No canonical tests or Python library files were modified, no new RustPython behavior tests were added, and no diagnostic filtering or expected-output special cases were introduced.

Validation

The exact production source passed:

• Default-feature VM cargo check --offline --locked -p rustpython-vm --lib.
• Strict VM cargo clippy --offline --locked -p rustpython-vm --lib -- -D warnings.
• Default CLI cargo build --offline --locked --bin rustpython.
• 832 fixed compile-only observations against the pinned CPython oracle, with complete raw warnings and exception objects retained. Returned code objects/ASTs were not executed.
• Five unchanged canonical methods through normal unittest fixtures/decorators: f-string string-part backslashes, backslash-before-double-bracket, warns-once, raw t-strings, and triple-quoted t-strings. All passed with no skips or expected failures.

An independent readonly candidate binary was sealed and its actual native and Python-module origins verified. Every gate checked source/tool/Lib bindings. Candidate source patch SHA256: 2a3f8ba850233eee80a0a2df4972363697d1707ae57e42a40d3cd4a1f958a7df (final follow-up delta against validated v8); binary SHA256: eb75c0a58986a48142d3ae7986823b89c520c6c8224da8ba925841cbb565b9a5.

All 832 observations agree on warning lists and exception selection/chaining, including f/t body, format-spec, brace-tokenizer locations and the added numeric/ordinary-text corrections. The final follow-up corrected exactly four ordinary exec bare-CR rows; the other 828 rows were unchanged.

Exact results and remaining boundaries

Full raw equality is 382/832. A further 382 rows differ only in existing global SyntaxError args/repr construction. The remaining 68 rows retain other diagnostic differences: rejected TYPE_COMMENT validator ranges (40), Single generic-error LF text (18), the explicit malformed-source fallback control (10). These are recorded, not normalized away. General ordinary-string diagnostics outside the reviewed final-line scope are not claimed globally equivalent.

Arbitrary custom warning-callback exception conversion, complete malformed-source parser recovery, global SyntaxError argument tuples, and real-file filename lookup are not rewritten or claimed fully equivalent. Full workspace/CAPI suites and entire Python test modules were not run in this bounded validation.

AI assistance: OpenAI Codex (model version unavailable); commits contain Assisted-by trailers.