Address remaining SSA review comments

[voltjia]

Summary

• Keep the root package API conservative by exposing the SSA jit entrypoint without adding root-level aot or lower exports.
• Normalize auto-tuning naming across tests and user-facing errors while preserving the AutoTuner type.
• Align CompileRequest and its construction sites with the established arrangement, application, tensors order.
• Add the requested registry spacing and keep the compiler entrypoint coverage focused on supported behavior.

Scope

This follows up the four unresolved review threads on #164. It does not address the Triton AOT multi-device/context issue, which is covered by #195.

Validation

• Built and installed the editable package in accelerator-dev/nvidia:latest with pip install --no-deps --force-reinstall -e ..
• ruff check (Ruff 0.15.21)
• ruff format --check (Ruff 0.15.21)
• python scripts/check_contributing_style.py
• Current affected tests after the requested test cleanup: 96 passed
• Full NVIDIA single-device suite from the production-code commit is shown below. The subsequent commits only adjust test coverage and import ordering. The multi-device case is skipped for Fix Triton AOT multi-context launches #195; the TileLang reload parameter is deselected because the NVIDIA image does not include the optional TileLang dependency.

pytest output:

$ CUDA_VISIBLE_DEVICES=0 PYTHONPATH=src python -m pytest \
    --deselect='tests/test_built_artifact_reload.py::test_aot_built_artifact_can_be_reloaded[tilelang-cuda]'

collected 338 items / 1 deselected / 337 selected

========== 336 passed, 1 skipped, 1 deselected in 1851.00s (0:30:50) ==========