fix(core): normalize GPT locate coordinates

[quanru]

Summary

• request GPT-5 bounding boxes in the existing normalized 0-1000 coordinate format
• map normalized boxes against Midscene's authoritative prepared screenshot dimensions
• retain strict out-of-bounds validation instead of clamping invalid model output
• add an edge-element regression covering the reported 1440 x 3200 screenshot

Why

GPT does not know the authoritative pixel dimensions of the image after provider-side image handling. Asking it for absolute pixels allowed a valid bottom-edge element to be returned beyond the local screenshot height. Normalizing the model contract removes that size guess and lets Midscene perform the only pixel conversion.

Fixes #2929

Validation

• pnpm exec nx test @midscene/core -- tests/unit-test/model-adapter/gpt.test.ts (11 tests)
• pnpm exec nx test @midscene/core -- tests/unit-test/prompt/prompt.test.ts (34 tests)
• pnpm exec nx build @midscene/core
• pnpm run lint
• git diff --check

No live GPT request was run because the required model credentials were not available; the adapter prompt and coordinate conversion are covered directly.