Keep GPU backtests and RunPod; export raw results without analysis

[howardc38]

The public package now runs GPU strategies and exports raw results, without imposing post-backtest analysis. Remove date splitting, common-parameter filtering, charts, effect-size statistics and Top-N ranking. Preserve GPU kernels, strategy plugins, the educational RSI example, CPU/GPU validation, performance comparison and RunPod lifecycle.

Each run writes four exact float64 grouped arrays to NPZ and a schema-2 manifest with ordered dimensions, counts, fees, shapes, reduction checks and a file hash. Single-combination output is supported. RunPod defaults to one run; removed analysis options fail before allocation. The historical 10.34× result remains identified as the v0.4 measurement.

Validation:

• Two review rounds across numerical correctness, dependency/contracts, maintainability and RunPod security; findings fixed.
• 96 local tests pass, including isolated CUDA simulation. Lint, format, source/wheel build and public-content checks pass.
• Clean wheel installed outside checkout; packaged example and RunPod bundle verified.
• Actual RTX 4090: v0.5/v0.6 all four arrays are byte-identical across four RSI workloads, including 16,777,216 and 1,000,000,000 pairs × 1,024 bars.
• Current NPZ matches API values; schema, dimension order, fees and SHA-256 checked. Default RunPod bundle output matches the clean wheel.
• All five physical-GPU tests pass, including independent single-combination profit/fee/loss cases.
• Owned pod deleted; API read-back confirms zero active pods.

Raw evidence: benchmarks/results/rtx4090_raw_output_parity_20261005.json.

Migration: ranked CSV/schema-1, analysis commands, top_n, date-split config and RunPod pipeline/--charts are removed. The numerical four-array return remains unchanged.