Adds no-topk variant for kernel performance analysis

[LoserCheems]

Introduces a new benchmark function that skips topk computation to isolate
pure kernel performance from selection overhead. This enables more accurate
measurement of the core attention mechanism by removing the topk bottleneck.

Expands benchmark configurations to include inference scenarios with varying
key lengths and updates output formatting to display both standard and
no-topk performance metrics with speedup comparisons.

Provides insights into topk overhead impact on overall performance.

[Copilot]

Pull Request Overview

This PR introduces a new benchmark function that isolates kernel performance by skipping the topk computation. It expands benchmark configurations to capture both standard and no-topk performance and updates output formatting to include additional speedup comparisons.

• Added a new function dynamic_mask_attention_cuda_no_topk to measure kernel performance without topk overhead.
• Updated benchmark loops and result reporting to include no-topk metrics.
• Extended the benchmark configuration with new inference scenarios.

Comments suppressed due to low confidence (1)

benchmarks/benchmark_forward_performance.py:248

• Consider adding an inline comment explaining why keep_window_size is set to 0 in the subsequent call to apply_dynamic_mask_attention, clarifying its role in skipping the topk computation for kernel performance measurements.

    zero_hold_states = calculate_zero_hold_states(value_states, dt_proj, A, causal_mask)

[Copilot]

[nitpick] The loops for warmup and benchmark runs for dynamic_mask_attention_no_topk are similar to those used for the standard dynamic_mask_attention. Consider refactoring these into a helper function to reduce duplication and improve maintainability.

[Copilot]

[nitpick] The header printing uses hard-coded column widths. Consider extracting these widths as constants or using a formatting helper to improve flexibility and maintainability.

Suggested change

	print(f"🔧 {'Configuration':<42} ⚡ {'Flash (ms)':<12} 🚀 {'DMA (ms)':<12} 🚀 {'DMA-Skip-All (ms)':<22} 📈 {'Speedup':<12} 📈 {'Skip-All-Speedup':<20} 💾 {'Memory':<10}")
	print("🔄" + "-" * 155 + "🔄")
	print(f"🔧 {'Configuration':<{CONFIGURATION_WIDTH}} ⚡ {'Flash (ms)':<{FLASH_WIDTH}} 🚀 {'DMA (ms)':<{DMA_WIDTH}} 🚀 {'DMA-Skip-All (ms)':<{DMA_SKIP_ALL_WIDTH}} 📈 {'Speedup':<{SPEEDUP_WIDTH}} 📈 {'Skip-All-Speedup':<{SKIP_ALL_SPEEDUP_WIDTH}} 💾 {'Memory':<{MEMORY_WIDTH}}")
	print("🔄" + "-" * (CONFIGURATION_WIDTH + FLASH_WIDTH + DMA_WIDTH + DMA_SKIP_ALL_WIDTH + SPEEDUP_WIDTH + SKIP_ALL_SPEEDUP_WIDTH + MEMORY_WIDTH + 20) + "🔄")

[SNHuan]

LGTM

[SNHuan]

LGTM