[xegpu] Clean kernel_bench compiler cache

[tkarna]

Minor fixes to xegpu kernel_bench routine

• Clean up execution engine cache after each kernel to avoid device out-of-memory errors.
• Simplify kernel execution logic.
• Add --input-shapes option to override kernel input tensor shapes.
• Add support for gpu-outlining stage kernel dump.
• Adds xegpu-wg dump CI test for KernelBench scaled attention kernel.