Linter: Fix byte-offset/UTF-16 index mismatch corrupting autofix and offense locations

[drjayvee]

Heads up: this commit and PR are LLM-generated and only superficially reviewed by me. The bug is real though! End of human output :D

Summary

Prism reports node locations as UTF-8 byte offsets, but several linter rules and the erb-prefer-direct-output autofix used those offsets directly as UTF-16 string-index offsets when calling source.substring(). This is only correct when the file is pure ASCII — any multi-byte character earlier in the file (even inside an unrelated tag or comment) shifts every subsequent byte offset, corrupting the result.

This is what caused reports like:

-            <%= "#{card_event.event_type}(#{data})" %>
+            <%= ard_event.event_type} %>(<%= ata} %>)

when running herb-lint --fix on a file that has any non-ASCII character (accented letters, emoji, etc.) earlier in the source.

• Adds stringIndexFromByteOffset to @herb-tools/core to correctly convert a Prism UTF-8 byte offset to a UTF-16 string index.
• Uses it in the erb-prefer-direct-output autofix (ERBStringToDirectOutputRewriter.extractStringContent / extractExpressionSource), which was corrupting the rewritten source.
• Uses it in locationFromOffset (shared by erb-no-instance-variables-in-partials, erb-no-debug-output, erb-no-unused-literals, erb-no-unused-expressions, erb-prefer-direct-output), which was reporting the wrong offense line/column whenever non-ASCII characters preceded the offense.
• Adds a substringFromByteOffset helper and uses it in the 3 rules that extracted a source snippet for their message text (erb-no-debug-output, erb-no-unused-literals, erb-no-unused-expressions), which were showing a corrupted snippet in the same scenario.

Test plan

• New autofix regression tests in erb-prefer-direct-output.autofix.test.ts reproducing the reported bug (preceded by an accented character / emoji)
• New unit tests for stringIndexFromByteOffset in core/test/prism/byte-offset.test.ts
• New regression tests for the location/message corruption in erb-no-debug-output.test.ts, erb-no-unused-literals.test.ts, erb-no-unused-expressions.test.ts, erb-no-instance-variables-in-partials.test.ts
• Verified end-to-end against real Prism-parsed byte offsets (via the published 0.10.2 packages, patched with this fix) that all of the above scenarios produce corrupted output before the fix and correct output after