perf: avoid full-frame promotion during alignment

[EricSpencer00]

What this changes

Fixes #2. Canonical alignment now promotes only the four sampled pixel neighborhoods to float64 instead of converting the entire source photograph before indexing. The existing interpolation, clamping, rounding, output dtype, and seeded measurement path are unchanged. An exact-output regression test compares the optimized sampler with the pre-change implementation at image edges and out-of-bounds coordinates.

Baseline and measured result

Baseline commit: f0ebfe39d9b2ef82e98343b3b24bcb794e9e5041 (the post-batching code before this change). On the same NUC runner, Python 3.12.3, NumPy 2.5.3, synthetic 1600x2000 RGB input, 512 crop, and 2048 samples, an alternating same-process A/B harness measured:

metric	old	new	change
direct sampler median	121.365 ms	96.813 ms	20.2% faster
alignment stage median	151.686 ms	89.648 ms	40.9% faster
full analysis median	389.171 ms	324.476 ms	16.6% faster

Both paths returned 42 measured and 26 unavailable results with no failures.

Verification

• Exact commit 86a7e5b: pytest -q -r s — 850 collected, 840 passed, 10 expected skips.
• Changed-file Ruff: passed.
• Changed-module mypy under Python 3.12: passed.
• python -m build and twine check dist/*: passed.
• Repository-wide Ruff still reports the pre-existing 83-error backlog; CI already treats that check as non-blocking.
• Independent Luna review was requested twice at gpt-5.6-luna/max, but the review transport was unavailable; no substitute model was used.

Risk and rollback

The code path is limited to the canonical crop sampler. The regression oracle preserves the previous pixel output exactly for boundary cases. Revert commit 86a7e5b to roll back.