ingest: bind PostgreSQL array literals wide; LFConversion=0 for psqlodbc on Windows

[singhpratech]

Fixes the five Windows-only failures in tests/test_pg_array_ingest.py found while verifying #107 on Windows (they are pre-existing there, not #107's: the same five fail against main).

Cause

Two things, both invisible on Linux and macOS:

1. The array literal was bound narrow. ArrayIngestExecuteBatch bound each column's PostgreSQL array literal as SQL_C_CHAR / SQL_LONGVARCHAR, while the row-at-a-time path binds a string parameter wide. A Unicode driver reads a narrow parameter in the client's ANSI code page, and on Windows psqlodbc re-encodes the UTF-8 bytes to UTF-8 a second time, so "héllo wörld" landed as "héllo wörld". On Linux and macOS the UTF-8 locale makes the same bind a pass-through.
2. psqlodbc's LFConversion defaults to 1 on Windows (0 on every other platform, DEFAULT_LFCONVERSION in its dlg_specific.h). At 1 the driver rewrites every LF in a fetched text value to CR LF: "line1\nline2" read back as "line1\r\nline2", bytes the server never stored (checked with octet_length/encode(...,'hex') on the tables: LF only). Not specific to the array path, it just happens to be the only test with a newline fixture.

Fix

• src/odbc_bind.c: the literal now travels the way BindRowParam sends a string: SQL_C_WCHAR / SQL_WLONGVARCHAR, converted with the existing UTF-8-to-SQLWCHAR helper (so the iODBC four-byte SQLWCHAR case is covered by wide_utf16_pairs as elsewhere), under the same wchar_as_utf8 / narrow_params condition, so drivers whose narrow path is UTF-8 keep it. One SQLWCHAR buffer per column, reused across executes like the narrow one.
• src/odbc_driver.c: the connection-keyword auto-tuning appends LFConversion=0 on Windows only, for psqlodbc (recognised by its Driver= value, or by UseDeclareFetch behind a DSN), when the caller has not set LFConversion. The caller's own value still wins, and adbc.odbc.tune=false turns it off with the rest. The rule that nothing changing a query's result is ever set stands: 0 is the value that returns what Linux and macOS already do. Comment block and docs/how-it-works/connection-keywords.md updated accordingly.

No C unit test fits (both halves need a live driver); tests/test_pg_array_ingest.py is the regression guard and now passes in full on Windows.

Verification

Windows 11 x64, MSVC 19.44 (Visual Studio 17 2022), Release; psqlodbc 18.00.0002 ("PostgreSQL Unicode(x64)"); PostgreSQL 16 in Docker; Python 3.12, adbc_driver_manager 1.12.0, pyarrow 25.0.1.

check	before (main)	after
ctest	7/7	7/7
tests/test_pg_array_ingest.py	13 passed, 5 failed	18 passed
tests/ (minus tests/validation, needs a package not installed here)	165 passed, 26 skipped, 6 failed	183 passed, 26 skipped, 6 failed
tests/compat/test_matrix.py postgres / sqlite / duckdb	PASS	PASS
python/tests	22 passed	22 passed
Verified on macOS (a7c9f18)	macOS 26.6.2 arm64, Apple clang 21.0.0, unixODBC 2.3.14, Debug build: ctest 10 / 10 (smoke skipped). PostgreSQL 16.15 (scratch cluster) via psqlodbc 18.00.0004: tests/test_pg_array_ingest.py 18 / 18 (also 18 / 18 on main there: the wide bind is a no-op on a UTF-8 locale, as expected). SQLiteODBC 0.99991: tests/test_sqlite.py 9 / 9. test_matrix.py postgres PASS, sqlite PASS. A local merge of #107 + #109 on main (not pushed) builds clean, ctest 10 / 10, tests/test_postgres.py + tests/test_pg_array_ingest.py 34 / 34.
Verified on Linux (db18685)	Ubuntu 24.04 x86_64, gcc, Debug build: ctest 10 / 10. PostgreSQL 16.15 (compat compose) via psqlodbc 16.00.0000: tests/test_pg_array_ingest.py 18 / 18; test_matrix.py postgres PASS, sqlite PASS. A local merge of #107 + #109 builds clean: ctest 10 / 10, tests/test_postgres.py + tests/test_pg_array_ingest.py 34 / 34. Second commit refreshes rust/csrc and narrows the Driver= match to psqlodbc's registered names.

The six tests/ failures are the same set on main and on this branch: four test_delegate.py cases (no native driver forwarding on Windows, IM002 wording) and the two test_plug_and_play.py install flows (multi-config cmake --install, install.sh under WSL). Unrelated to this change.

Linux and macOS still need a run of tests/test_pg_array_ingest.py and tests/test_postgres.py: the wide bind is the only half that changes behaviour there (the LFConversion half is #if defined(_WIN32)), and it should be a no-op since the row path already binds the same way.