Add performance regression tests for cuda.core - #2

Description

Adds automated performance regression tests for the cuda.core API to detect when changes inadvertently slow down critical CUDA operations.

Changes:

• New test_performance_regression.py with two tests:
  • Kernel launch overhead benchmark (1000 empty kernel launches)
  • Memory transfer bandwidth benchmark (1KB, 1MB, 64MB transfers)
• Register performance pytest marker in conftest.py
• CI workflow updated to run performance tests on A100 hardware only
• PERFORMANCE_THRESHOLDS.md documenting threshold values and update procedures

Thresholds:

• Kernel launch: < 50μs average
• Memory bandwidth: 0.1/5/10 GB/s for 1KB/1MB/64MB respectively

Human Review Checklist

• Verify device.allocate() API: Confirmed - Device.allocate(size, stream) exists at line 1330 in _device.pyx and delegates to self.memory_resource.allocate(size, stream).
• Threshold values: The thresholds are estimates based on A100 specs. May need adjustment after running on actual CI hardware.
• Kernel launch test lacks warm-up: Consider if a warm-up iteration is needed before timing (though 1000 iterations should average out cold-start effects). Note: memory transfer test does include warm-up.
• CI validation: Tests have not been validated in CI yet (forked repo may not have workflows enabled). Verify tests pass on actual A100 hardware.

Checklist

• New or existing tests cover these changes.
• The documentation is up to date with these changes.

Link to Devin run: https://app.devin.ai/sessions/2db856c666e84ea98b2800312a6f0ce3
Requested by: Shawn Azman ([email protected]) / @ShawnAzman