Yield explicitly supplied headers only once when reading incrementally

[OskarEichler]

Summary

Clear the pending raw headers before yielding their synthetic row. This lets later shift/each calls advance into data and reach EOF, while rewind rebuilds the parser.

Reproduction

csv = CSV.new("1,2\n", headers: %w[a b], return_headers: true); 3.times { p csv.shift } returns the header three times. This became observable when #363 removed the persistent parser enumerator.

Verification

108 focused checks: 72 failures on the release, zero with the patch, across array/string/empty headers, converters, unconverted fields, incremental reads, repeated EOF, and rewind.

This isolated patch passes the unchanged upstream suite in both normal and CSV_PARSER_SCANNER_TEST=yes modes: 527 tests / 4,050 assertions in each mode, zero failures/errors. Commands use Ruby 4.0.6 through rbenv (bundle exec ruby run-test.rb). Syntax and git diff --check pass. The combined release-based consumer branch also passes gem build and rake warning:error rdoc.

No tests were added or changed, per the requesting project's explicit no-new-tests policy. Focused checks and measurements live outside this repository. The upstream no-RuboCop policy is respected. Verification was local on macOS/Ruby 4.0.6; the supported OS/Ruby matrix was not run locally. Current main and installed 3.3.6 share base commit 0873ab362d996f12796c3c3e8998b5be657a9b12; no unrelated changes are included.

Breaking changes and limitations

None intended. Repeated incremental reads no longer emit an endless stream of synthetic headers.