trunk/40b4ed1a18c5c250518e50fbcccc213c61e2c2cc: [aten] Fix CPU and MPS logit for eps > 0.5 (#181297)

Issue

Fixes #177839.

Summary

For eps > 0.5 the clamp bounds invert (lo = eps > hi = 1 - eps). The scalar and CUDA kernels resolve this as x < lo ? lo : (x > hi ? hi : x), so lo wins. Two paths disagreed and produced a sign-flipped result:

• CPU vectorized: vec::clamp = min(hi, max(lo, x)), so hi won.
• MPS logit: clampWithTensor, also min(max(x, lo), hi), so hi won.

Both now apply lo last (nested blendv on CPU, nested select on MPS), matching scalar and CUDA. CUDA already conformed, so all three backends agree. eps > 0.5 is mathematically ill-defined; the goal is cross-backend consistency.

Tests

Added test_logit_vectorized_matches_scalar (eps {0.49, 0.51, 0.6, 0.9} x 4 dtypes, bit-exact vs the scalar kernel) and an eps=0.6 sample_inputs_logit sample so test_ops.py and the MPS consistency tests cover eps > 0.5. Verified on Apple Silicon: test_mps.py -k logit, test_ops.py -k logit, and test_unary_ufuncs.py -k logit pass, including the previously failing test_output_grad_match_logit_mps_float32.

BC-breaking?

No.

Pull Request resolved: #181297
Approved by: https://github.com/jansel