fix(python): reduce Windows release wheel size (#4241)

The Windows wheel for 0.39.0 exceeded PyPI's 100 MiB file-size limit,
preventing the original release upload. Restore the repository's release
profile (`fat` LTO, 1 codegen unit) by removing the Windows-only
`thin`/16 overrides introduced in #3716. Keep `rust-lld` as the linker.

This recovers wheel-size headroom at the cost of the longer fat-LTO
build. The already-published 0.39.0 Windows wheel was recovered
separately by recompressing the original artifact; this change addresses
the build configuration for future releases.

Fixes #4239.

### Windows size comparison

Built the same v0.38.0 source
(`8c68e0c619f2b1febe92e51a29d72c968d24a5c9`) twice on one AWS
`m7i.4xlarge` Windows Server 2025 machine, changing only the
LTO/codegen-unit overrides:

| Artifact | thin LTO / 16 units | fat LTO / 1 unit |
| --- | ---: | ---: |
| Installable Windows wheel | 104,129,447 bytes (99.31 MiB) | 73,875,399
bytes (70.45 MiB) |
| Uncompressed native module | 309,969,408 bytes (295.61 MiB) |
206,609,408 bytes (197.04 MiB) |

The thin/16 configuration increased wheel size by **40.95%** and
native-module size by **50.03%** relative to fat/1. The compressed
native module accounts for all but 3 bytes of the wheel increase.

Both builds used Rust 1.97.0, maturin 1.12.4, Python 3.13.5, MSVC
14.44.35207, Windows SDK 10.0.26100.0, rust-lld, static CRT, default
features, and `maturin build --release --strip --locked --verbose`. They
ran sequentially with separate empty target directories and identical
locked third-party dependencies. The tag's stale workspace package
versions were normalized once before both builds. This controlled pair
used VS2022 Build Tools; the historical GitHub runner used VS2026.

Both wheels passed ZIP/RECORD integrity checks and installed
successfully. Native smoke checks covered import, database creation, row
count, nearest-vector query, and reopening the database. This
establishes the combined configuration effect; it does not isolate LTO
mode from codegen-unit count or establish a runtime-performance
difference.

The figures above are the v0.38.0 reproduction, **not measurements of
this PR head**. The existing PyPI Publish pull-request workflow rebuilds
the current revision without publishing; its result is pending.

Workflow validation passed with `actionlint -shellcheck=`. Full
actionlint reports the same three pre-existing ShellCheck diagnostics in
the unchanged repository-selection step as on `main`.

.github/workflows/pypi-publish.yml

@@ -129,12 +129,8 @@ jobs:
       # link.exe is single-threaded and the long pole on Windows builds. Use
       # rustc's bundled lld-link instead.
       CARGO_TARGET_X86_64_PC_WINDOWS_MSVC_LINKER: rust-lld
-      # Fat LTO of the cdylib is single-threaded and the peak-memory step of the
-      # build. ThinLTO parallelizes it across the runner's cores, at some cost
-      # to runtime performance on our least performance-sensitive platform.
-      # Matches what the nodejs Windows builds already do in npm-publish.yml.
-      CARGO_PROFILE_RELEASE_LTO: thin
-      CARGO_PROFILE_RELEASE_CODEGEN_UNITS: 16
+      # Keep the repository's fat LTO / 1 codegen unit release profile to leave
+      # headroom below PyPI's file-size limit for Windows wheels.
     steps:
       - uses: actions/checkout@v6
         with: