ci(python): add opt-in native Intel macOS wheel validation - #4140

Summary

Add an opt-in native Intel macOS wheel workflow using the explicit macos-15-intel runner. This respects the low-volume/platform-cost rationale in #2836: it does not restore Intel to the default release matrix or publish anything to PyPI.

Downstream users still run local embedded LanceDB on Intel Macs. This provides a source-build path with an installed-wheel smoke test.

Changes

• Manual dispatch, plus PR validation only when this workflow or its smoke script changes.
• Existing macOS composite build action, repository release profile, locked dependencies, native x86_64 target and default feature set including fp16 kernels.
• Fresh venv outside the source tree; installed-version and Mach-O architecture checks.
• Actual create/insert, additive column migration, filtered vector search, update, reopen and delete smoke tests.
• Short-lived wheel artifact with source commit, Cargo.lock, Rust version and SHA-256 checksum; documentation states the compatibility/acceptance boundaries.

Validation

• Native Intel build and installed-wheel smoke passed for this PR's source at fa824b927d2fa40072cba1472e625ce9df746434: run 34060681299.
• Artifact: lancedb-0.39.0b3-cp310-abi3-macosx_10_15_x86_64.whl; downloaded artifact independently SHA-256 checked: c68028e55927e5cba9703c4c1dbbb176b16d589afc7de6b8633e3cc433f8ebba.
• Fresh installed environment passed pip check, native x86_64 Mach-O validation, insert/additive schema/filtered ANN/update/reopen/delete checks. No source-tree import was used.
• The only follow-up source change raises the workflow budget from 120 to 240 minutes to match the successful validation run; build flags and smoke code are unchanged. No LTO/release-profile relaxation.
• Ruff lint and formatting passed for the smoke script; whitespace check passes.
• Separate downstream pinned LanceDB 0.34.0 verification also passed in run 34048642479; this is distinct from the 0.39.0b3 evidence above.

A 10.15 deployment tag is not a claim that Catalina or all older Intel CPUs were tested. Downstream ML dependencies need their own validation. No production data, credentials, publish permissions or release changes are involved.