Use block-scoped atomics in BlockHistogramAtomic on SM 60+

[NitishNaineni]

Closes #3357

Summary

• Replaces atomicAdd with atomicAdd_block in BlockHistogramAtomic::Composite on SM 60+ using NV_IF_TARGET, with fallback to atomicAdd on older architectures.
• This matches the pattern already used in agent_histogram.cuh (lines 312, 325, 350).

Codegen verification

SASS confirms the scope change on SM 86 (RTX 3080):

• Before: RED.E.ADD.STRONG.GPU
• After: RED.E.ADD.STRONG.SM

Performance was identical on Ampere in my benchmarks (1024 threads × 16 items, per-block global histograms, various bin counts and grid sizes).

Test plan

• cub.test.block.histogram.bins_32 — all 1512 assertions passed
• cub.test.block.histogram.bins_256 — all 1512 assertions passed
• cub.test.block.histogram.bins_1024 — all 1512 assertions passed

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[miscco]

We only suppport SM75 and up, so this should just be

Suggested change

	NV_IF_TARGET(NV_PROVIDES_SM_60,
	(atomicAdd_block(histogram + items[i], 1);),
	(atomicAdd(histogram + items[i], 1);));
	atomicAdd_block(histogram + items[i], 1);

[bernhardmgruber]

We do accept PRs that add/maintain support for older architectures though. So this change is fine. But we are also fine with just the version that @miscco suggested.

@NitishNaineni I leave the choice to you whether you maintain the fallback for below SM60 or not.