fix(miniwob): reject partial find-greatest rewards

[jstar0]

Summary

• add a task-level MiniWoB success threshold while preserving the existing default
• require miniwob.find-greatest to clear its partial wrong-answer reward
• cover the behavior with unit-level and real-browser regressions

Problem

MiniWoB++ returns a raw reward of 0.1 when find-greatest submits the wrong
card and 1.0 for the correct card. BrowserGym currently converts every
positive raw reward to 1.0, so the wrong-card partial reward is reported as a
successful benchmark episode.

Changes

AbstractMiniwobTask now exposes a success_threshold that defaults to 0.0,
which preserves existing behavior for all other MiniWoB tasks.
FindGreatestTask overrides the threshold to 0.5, separating its known
0.1 wrong-card reward from its 1.0 success reward.

The regression coverage checks both the threshold conversion directly and a
real find-greatest browser episode that deliberately submits the minimum
card.

Scope and risk

This does not change dependencies, task registration, or the reward behavior
of tasks that do not override the threshold. The only task-specific behavior
change is that find-greatest no longer promotes its documented partial
wrong-answer reward to success.

Verification

uv run black . --check
uv run pytest tests/miniwob/test_base.py -k success_threshold
MINIWOB_URL=file://<miniwob-plusplus>/miniwob/html/miniwob/ uv run pytest tests/miniwob

Results: Black left all 91 files unchanged; the focused regression passed 3
tests; the complete MiniWoB suite passed 43 tests with 4 skipped.

Fixes #392