Separate GPU core from examples, tests and execution tools

[howardc38]

The runtime package mixed numerical engine code with RSI examples, CPU comparison code, hardware checks, and RunPod lifecycle logic, making it hard to see what the engine actually owns.

Separate the code by responsibility:

• src/gpu_backtest/core/: GPU runner/kernels, grid, data validation, indicators, statistics and output. No example/tool/workflow imports.
• src/gpu_backtest/workflows/ and cli/: generic analysis flows and thin command adapters.
• examples/gpu_backtest_examples/: the RSI plugin, its config/data and execution rules.
• tools/gpu_backtest_tools/: benchmarks/CPU baseline, checks/reference, and RunPod API/SSH/bundle/lifecycle helpers.
• tests/cpu/ and tests/gpu/: separate execution surfaces. Active docs cover three topics; testing instructions and historical evidence live with their owners.

Preserve the public run API, CLI names and old RSI shorthand at the API edge. Direct internal imports relocate under the new namespaces. Wheel/remote exports include the three explicit packages with example data and helper requirements.

Validation: 89 tests passed; lint/format/build/release-content checks passed. GPU kernel AST is identical to the previous version. A v0.4 wheel versus v0.5 simulator comparison produced byte-identical grouped arrays, both CSVs and manifest. A fresh wheel ran the example, pipeline and charts; RunPod dry-run/export was built successfully without allocating resources. Historical benchmark JSONs are unchanged. No numerical GPU logic or private repository files were edited.