Make Triton benchmark roofline GPU-aware

[peter941221]

This removes a hard-coded RTX 4090 assumption from the Triton decode-attention benchmark.

Failure mechanism:

• benchmarks/benchmark_triton.py hard-coded RTX 4090 roofline numbers into roofline_analysis()
• on other GPUs, especially Blackwell-class consumer GPUs like RTX 5090, the reported ridge point and bandwidth-bound latency are inaccurate even when the benchmark itself runs correctly
• that weakens the benchmark's explanatory value for anyone validating kernels on newer hardware

Semantic change:

• add known roofline presets for RTX 4090 and RTX 5090
• resolve the preset from torch.cuda.get_device_name(0)
• allow manual CLI overrides for FP16 TFLOPS and bandwidth when running on an unrecognized GPU or using a custom assumption set
• add a CPU-safe test covering preset selection, override behavior, and roofline-analysis output wiring

Preserved invariant:

• benchmark execution flow is unchanged
• unknown GPUs still run, now with an explicit fallback label instead of silently implying RTX 4090 numbers are universal

Testing:

• pytest tests/test_benchmark_triton_roofline.py -q