[xiao115255]

Every action grounds the same point twice. Worker.generate_next_action checks the model's
response with CODE_VALID_FORMATTER, which calls create_pyautogui_code and therefore evaluates
the action: agent.click("...") reaches OSWorldACI.generate_coords and calls the grounding model.
The worker then evaluates the same code again to run it, so generate_coords runs a second time
with the same screenshot and the same referring expression. The model is asked for the same point
twice and only the second answer is used.

generate_coords now keeps the coordinates it has read for the current screenshot, keyed by
referring expression, and drops them whenever the screenshot changes. Coordinates can't outlive
the screen they were read from, and the second call is free. An equal-but-distinct screenshot just
misses the cache, so a miss can never return a stale point.

This is s3 only: CODE_VALID_FORMATTER exists nowhere else, so the s2/s2_5 copies of
generate_coords don't need it.

Fixes #178

How I tested it: tests/test_grounding_coords_cache.py patches the grounding call and counts it
for the two-call sequence above — one call with this change, two without it — plus a new screen and
a second element on the same screen, to check the cache doesn't cross wires. On the unpatched
grounding.py those two cache assertions fail and the rest pass, so the test does fail without the
change. pytest tests → 9 passed; black --check gui_agents → unchanged. Run on Windows with
Python 3.13.