Perf: autotune emits three cudaMemsetAsync per chunk-mult combination

[felixx-sp]

Confidence: medium · Effort: trivial (<1 h)

Problem

The autotune loop resets d_keys_tested, d_best_packed, d_stop_requested between every tpb/bpsm/chunk_mult combination with three separate cudaMemsetAsync calls followed by a blocking cudaStreamSynchronize. Three tiny resets cost more in launch overhead than the bytes moved.

Files: src/bruteforce.cu:2150-2153

Suggested fix

Either:

1. Allocate a single 24-byte device counter struct and reset it in one cudaMemsetAsync, or
2. Batch into a small 1-thread kernel that zeroes all three.

Why it matters

Cuts ~2 ms × N_combinations off autotune startup. Not huge but a clean, low-risk win.