perf: GPU historical scan 2x slower than v1 (162s vs 83s)

[cfournel]

GPU_ScanHistoricalCandles now takes ~162s for 2022-2026 vs 83s in the original commit 852afd6. CLBuffer API param order changed — may have introduced data transfer overhead or the GPU scan is less effective. Profile and fix.

[cfournel]

False alarm — no GPU regression. New API is actually 2.3× faster.

The 83s→146s comparison in issue #45 was comparing apples to oranges:

• Old baseline (backtest_baseline.json, wall_time_s=83): 2025-only data (52M ticks)
• New 2022-2026 run (146s): 160M ticks — 3× more ticks, naturally takes longer

Correct comparison (same 2025-only period):

Version	Wall time	Ticks
Old GPU (852afd6, old KZ, 17T)	83s	~52M
**New GPU (new MT5 API, current)	36s	52M**

The new API is 2.3× faster for the same data period. The MT5 API update did not regress GPU performance — if anything, it improved it.

GPU_ScanHistoricalCandles function: Only scans inpHistoricalScanBars=30 bars at init — negligible overhead. The total wall_time_s measures the full backtest tick loop, not just the GPU init scan.