fix: round 2 backtest engine correctness and performance fixes

[Copilot]

Second optimization pass on the EasyQuant backtest framework, targeting correctness bugs in order execution and attribution, plus memory/performance improvements.

Bug fixes

• ORDER_TARGET partial fill logic (engine.py): Remaining order after partial fill now keeps the original target_amount. The engine already computes delta = target - current_position on the next bar, so subtracting filled shares caused systematic undershoot.

# Before (wrong): target 1000, filled 300 → remaining target = 700
# Next bar: delta = 700 - 300 = 400 (misses 300 shares)
remaining_req["target_amount"] = target - filled

# After (correct): target 1000, filled 300 → remaining target = 1000  
# Next bar: delta = 1000 - 300 = 700 (correct)
remaining_req["target_amount"] = order_req["target_amount"]

• ORDER_TIMEOUT crashes attribution (attribution.py): _calc_trade_win_rate() and _calc_profit_loss_ratio() now skip non-BUY/SELL entries. ORDER_TIMEOUT log entries lack price/amount keys, causing KeyError.

• NaN propagation in excess metrics (attribution.py): _calc_excess_metrics() applies fillna(0) on benchmark returns before subtraction, preventing NaN from poisoning all downstream excess return metrics.

Performance / memory

• Unbounded _file_locks dict (data_cache.py): Capped at 1024 entries with eviction of unlocked entries. Previously grew monotonically in long-running services.

• _to_tv_date() simplification (report.py): Collapsed 5 isinstance checks to 1 — datetime.date covers datetime.datetime and pd.Timestamp via inheritance.

• _bare_code() helper (engine.py): Extracted the repeated .replace(".XSHG","").replace(".XSHE","") pattern (6 call sites) into a single function.

Tests

17 new tests in test_round2_fixes.py covering all fixes. 391 total pass (374 + 17).