chore: Release 3.17.1 (#3436)

* fix(lambda): exclude auditwheel-vendored libs from strip to preserve ELF alignment

GNU strip on auditwheel's vendored libraries (numpy.libs/libscipy_openblas64_.so
and friends) rewrites their LOAD-segment file offsets without preserving the ELF
invariant p_offset == p_vaddr (mod p_align). mmap-based loaders tolerate the
violation, but Lambda's loader rejects it at import time with:

    ELF load command address/offset not page-aligned

This broke 'import numpy' on Lambda for py3.12-3.14 (both x86_64 and arm64) once
the layer switched to PyPI manylinux wheels. Exclude *.libs/ from the strip find;
our own extension modules are still stripped.

* docs: update Lambda layer ARNs for 3.17.1

Author: kukushking

building/lambda/build-lambda-layer.sh

@@ -117,7 +117,14 @@ rm -rf python/s3transfer*
 cp -r /aws-sdk-pandas/dist/pyarrow_files/pyarrow* python/
 
 # Removing nonessential files
-find python -name '*.so' -type f -exec strip "{}" \;
+# Skip auditwheel-vendored libraries (numpy.libs/, scipy.libs/, ...): these are
+# pre-built shared objects (e.g. libscipy_openblas64_.so) whose LOAD segments use
+# large-page alignment. GNU strip rewrites their file offsets without preserving
+# the ELF invariant p_offset ≡ p_vaddr (mod p_align), which mmap-based loaders
+# tolerate but Lambda's loader rejects with "ELF load command address/offset not
+# page-aligned" (issue: numpy C-extension import fails on Lambda). Our own
+# extension modules (awswrangler/pandas/pyarrow) are safe to strip.
+find python -name '*.so' -type f -not -path '*.libs/*' -exec strip "{}" \;
 find python -wholename "*/tests/*" -type f -delete
 find python -regex '^.*\(__pycache__\|\.py[co]\)$' -delete