Deduplicate Max/Min reduction gradient and add Min forward-mode

[cetagostini]

Description

This builds on the recently-merged #2253 / dc503f117 ("Add gradient (pullback) to Min Op"), which fixed the missing Min gradient by copy-pasting Max.pullback into Min. This PR:

1. Deduplicates that gradient logic into a shared base class MaxAndMinCAReduce(NonZeroDimsCAReduce), so Max and Min cannot drift apart. The reduction subgradient is identical for both — the cotangent routes to the position(s) where the input equals the reduced output — so a single source of truth is the natural design. Max/Min become thin subclasses binding only their scalar op.

2. Adds the forward-mode (pushforward / R_op) that Min still lacks, and generalizes Max's. The old Max.pushforward was matrix-only (raised NotImplementedError for >2D or multi-axis inputs) and — via its Argmax-based single-winner logic — was not the adjoint of its own pullback at ties. The new form

   out_dot = (eq(out_pad, x) * x_dot).sum(axis=axis)

   is the exact transpose (adjoint) of the pullback, supports arbitrary ndim/axis, and gives Min forward-mode for free. The adjoint identity ⟨J·ẋ, ḡ⟩ = ⟨ẋ, Jᵀ·ḡ⟩ now holds on ties (it previously did not).

Tests

• TestPushforwardPullback.test_min — forward/reverse-mode parity with test_max, exercising the bare Min op (which pt.min never instantiates, since it lowers to -max(-x)).
• TestPushforwardPullback.test_max_min_pushforward_on_ties — deterministic tied-input test pinning the tie-summing convention and the adjoint identity (random data is tie-free almost surely, so it would not guard this).

The reverse-mode regression is already covered by TestMinMax::test_grad_Min (added in #2253), which now runs through the shared base class. Existing Max/Min/CAReduce/uncanonicalize suites pass; gradients verified on the C, Numba, JAX, and MLX backends. Op identity (equality, hashing, pickling, __props__, __str__) is preserved.

Checklist

• Explain motivation and context (above)
• New tests added (forward-mode + tie adjoint-consistency)
• Pre-commit / ruff clean

🤖 Generated with Claude Code