[TEST] Cover Gluon Blackwell tcgen05 scaled ops

[Jokeren]

Summary

Add CPU-result coverage for Blackwell tcgen05_mma_scaled simulation through Gluon.

This adds a focused end-to-end test that:

• stores A/B tiles in shared memory
• stores uint8 scale groups in tensor memory scale layouts
• runs blackwell.tcgen05_mma_scaled with e4m3 format constexpr operands
• checks the simulated CPU result against a PyTorch scaled matmul reference

The branch also teaches the Gluon CPU runner to recover constexpr annotations from wrapped Gluon JIT functions and to wrap string constexpr arguments as tl.constexpr, which current Blackwell scaled MMA builtins require for format operands.

Validation

• PYTHONPATH=/mnt/keren/triton-viz pytest tests/end_to_end/test_gluon_blackwell_tcgen05_scaled_ops.py -q
• PYTHONPATH=/mnt/keren/triton-viz pytest tests/end_to_end/test_gluon_blackwell_tcgen05_scaled_ops.py tests/end_to_end/test_gluon_blackwell_tcgen05_ops.py tests/end_to_end/test_gluon_blackwell_tensor_memory_ops.py tests/end_to_end/test_gluon_wgmma_ops.py tests/end_to_end/test_gluon_tma_im2col_ops.py tests/end_to_end/test_gluon_blackwell_tma_ops.py tests/end_to_end/test_gluon_tma_ops.py tests/end_to_end/test_gluon_async_copy_ops.py tests/end_to_end/test_gluon_core_ops.py tests/end_to_end/test_gluon.py tests/end_to_end/test_core.py tests/unit/test_adapters.py -q
• pre-commit run --files triton_viz/core/simulation/gluon.py tests/end_to_end/test_gluon_blackwell_tcgen05_scaled_ops.py

[github-actions]

Performance Benchmark

Benchmark	main (min)	PR (min)	Change	Samples
gemm	0.105s	0.106s	+0.3%	20 / 20
gemm_oob	0.117s	0.117s	+0.1%	20 / 20
indirect_load	0.022s	0.022s	-1.3%	20 / 20
nested_loop	0.234s	0.236s	+1.0%	20 / 20
block_pointer_loop_advance	0.205s	0.205s	-0.2%	20 / 20
liger_jsd	0.137s	0.138s	+0.6%	20 / 20
flaggems_layernorm	0.398s	0.399s	+0.4%	20 / 20
swiglu	0.172s	0.172s	+0.2%	20 / 20
cross_entropy	0.984s	0.990s	+0.6%	20 / 20
fused_linear_jsd	0.210s	0.210s	+0.2%	20 / 20
Total	2.584s	2.596s	+0.4%	N/A

Iterations: 1 warmup + 20 measured
Samples are shown as main / PR; long pytest benchmarks may use fewer samples.