[7.1] Fix sequential Unicode decoding on big-endian platforms - #4961

Description

Backports #4945 from source commit 8dae63509ca17fa5b345596f7f3752bea7a4defd to release/7.1 in commit 50e099bd26558204b7a02f6705f69868e8a1ff53.

Fixes SqlDataReader.GetTextReader() with CommandBehavior.SequentialAccess corrupting nvarchar text on big-endian platforms. The decoder now reads only consumed UTF-16LE bytes without modifying the source buffer: a bulk span copy on little-endian hosts and explicit little-endian code-unit reads on big-endian hosts. Raw code units, including isolated surrogates, offsets, partial consumption, and zero-capacity conversions are preserved.

Release-specific assessment: 7.1 already contains the unified implementation, required database-object fixtures, and net462/net8.0/net9.0/net10.0 test targets. The implementation and both test files apply unchanged from the source commit; the TESTGUIDE insertion is adapted to the release document context. No dependency, framework, project-structure, public API, localization, or unrelated globalization changes are included. There is only one tracked implementation requiring the fix. The remote release base was refreshed before publication and remains 01126ba62, with no intervening conflicting fix.

• Tests added or updated
• Public API changes documented (not applicable; no public API changes)
• Verified against customer repro (real big-endian client unavailable)
• Ensure no breaking changes introduced

Issues

Backport of #4945. Preserves source provenance and contributor trailers; does not close the original merged PR.

Testing

Added 11 decoder regression cases covering source ownership, raw UTF-16LE code units, offsets, limited/zero capacity, odd trailing bytes, flush behavior, repeated partial reads, and empty conversion. Extended the four sync/async and buffered/streamed parameter integration cases with single-character and mixed-size reads, destination sentinels, zero-length reads, repeated EOF, and packet-spanning text. TESTGUIDE documents targeted execution and the architecture limitation.

Commands run from the repository root on Windows x64:

# PASS: 11 concrete decoder cases discovered.
dotnet test src\Microsoft.Data.SqlClient\tests\UnitTests\Microsoft.Data.SqlClient.UnitTests.csproj --framework net8.0 --list-tests --filter "FullyQualifiedName~SqlUnicodeEncodingTests" --verbosity quiet

# PASS: 11/11, zero skipped.
dotnet test src\Microsoft.Data.SqlClient\tests\UnitTests\Microsoft.Data.SqlClient.UnitTests.csproj --framework net8.0 --no-build --filter "FullyQualifiedName~SqlUnicodeEncodingTests" --blame-hang-timeout 10m --verbosity quiet

# PASS: 11/11, zero skipped.
dotnet test src\Microsoft.Data.SqlClient\tests\UnitTests\Microsoft.Data.SqlClient.UnitTests.csproj --framework net9.0 --filter "FullyQualifiedName~SqlUnicodeEncodingTests" --blame-hang-timeout 10m --verbosity quiet

# PASS: 11/11, zero skipped.
dotnet test src\Microsoft.Data.SqlClient\tests\UnitTests\Microsoft.Data.SqlClient.UnitTests.csproj --framework net10.0 --filter "FullyQualifiedName~SqlUnicodeEncodingTests" --blame-hang-timeout 10m --verbosity quiet

# PASS: 11/11, zero skipped.
dotnet test src\Microsoft.Data.SqlClient\tests\UnitTests\Microsoft.Data.SqlClient.UnitTests.csproj --framework net462 --filter "FullyQualifiedName~SqlUnicodeEncodingTests" --blame-hang-timeout 10m --verbosity quiet

# PASS: builds and discovers the conditional integration method.
dotnet test src\Microsoft.Data.SqlClient\tests\ManualTests\Microsoft.Data.SqlClient.ManualTests.csproj --framework net8.0 --list-tests --filter "FullyQualifiedName~UnicodeParameterAndReaderRoundTrip_PreservesWorldwideText" --verbosity quiet

The initial integration run without a working configuration skipped the conditional method (AreConnStringsSetup, TargetInvocationException); this was not counted as a pass. Subsequent live SQL Server validation used a temporary, untracked JSON configuration selected with the branch-supported $env:TEST_MDS_CONFIG variable. It used localhost SQL Server, tempdb, integrated authentication, Encrypt=false, and native SNI. Only the test process configuration changed; no server-wide or tracked configuration was modified. The existing RAII fixture cleans up transient tables. The scratch JSON was deleted and its absence verified afterward.

With that process-local override set:

# PASS: all four cases executed, zero skipped (3 s).
dotnet test src\Microsoft.Data.SqlClient\tests\ManualTests\Microsoft.Data.SqlClient.ManualTests.csproj --framework net8.0 --no-build --filter "FullyQualifiedName~UnicodeParameterAndReaderRoundTrip_PreservesWorldwideText" --blame-hang-timeout 10m --verbosity quiet

# PASS: all four cases executed, zero skipped (3 s).
dotnet test src\Microsoft.Data.SqlClient\tests\ManualTests\Microsoft.Data.SqlClient.ManualTests.csproj --framework net9.0 --filter "FullyQualifiedName~UnicodeParameterAndReaderRoundTrip_PreservesWorldwideText" --blame-hang-timeout 10m --verbosity quiet

# PASS: all four cases executed, zero skipped (3 s).
dotnet test src\Microsoft.Data.SqlClient\tests\ManualTests\Microsoft.Data.SqlClient.ManualTests.csproj --framework net10.0 --filter "FullyQualifiedName~UnicodeParameterAndReaderRoundTrip_PreservesWorldwideText" --blame-hang-timeout 10m --verbosity quiet

# PASS: all four cases executed, zero skipped (4 s).
dotnet test src\Microsoft.Data.SqlClient\tests\ManualTests\Microsoft.Data.SqlClient.ManualTests.csproj --framework net462 --filter "FullyQualifiedName~UnicodeParameterAndReaderRoundTrip_PreservesWorldwideText" --blame-hang-timeout 10m --verbosity quiet

# PASS before committing.
git diff --cached --check

Total: 44 decoder cases and 16 live SQL integration cases passed across the four test frameworks. All execution was on x64, which validates the little-endian path only. Real big-endian/s390x client execution remains outstanding; no claim of big-endian runtime validation is made.

Guidelines

Please review the contribution guidelines before submitting a pull request:

• Contributing
• Code of Conduct
• Best Practices
• Coding Style
• Review Process