[xegpu] Add torch llama3 attention example

[tkarna]

• Adds a KernelBench-like llama-bench_1B_attention.py kernel file which defines a Grouped-Query Attention (GQA) kernel following Llama-3.2-1B model specification.
• The kernel is executed with kernel_bench.py routine via the new -i/--input flag.
• The attention lowering schedules are updated to correctly lower the GQA layer. More IR canonicalization is required to fold 4D/5D expand/collapse_shape ops.
• In transform_to_flashattention the reduction_tile argument is now a (zero-padded) vector to allow easier handling of inferred tile sizes.
• Adds expand_output_destination transform op which rewrites materialize_in_destination(collapse_shape(X), D) into materialize_in_destination(X, memref.expand_shape(D)). This is required to avoid materializing a copy buffer for the layer result.
• Minor fix in contraction_normalization.