PR summary

Add a fast path for np.diff when n is at least the effective axis length. In that case the result is empty, so numeric and non-object dtypes can avoid repeated differencing of already-empty arrays. Object arrays keep the existing path so Python-level subtraction behavior is preserved.

ASV quick comparison against main:

bench_function_base.Diff.time_large_n((10,), 10):       21.0 us -> 16.5 us
bench_function_base.Diff.time_large_n((10, 10), 10):    41.3 us -> 13.9 us
bench_function_base.Diff.time_large_n((10,), 1000):      346 us -> 12.4 us
bench_function_base.Diff.time_large_n((10, 10), 1000):   481 us -> 15.0 us

First time committer introduction

I am a NumPy user/contributor looking at small, focused performance improvements in common array operations. I will handle review discussion and follow-up changes directly.

AI Disclosure

LLM / Harness: hybrid.

AI assistance was used to inspect the codebase, run local commands and benchmarks, draft PR text, and help prepare candidate patches. I reviewed the generated suggestions, selected the final changes, and am responsible for the code, submission, and follow-up discussion. Some PR text and candidate code edits were drafted with AI assistance; the final patch and text were reviewed and edited by me.