Fix pad with an axes subset and negative axes

[kapellirohith]

Proposed changes

edge_pad ignores its axes argument and indexes the pad-size vectors by
array axis rather than by position in axes (ops.cpp:1550, 1556). Padding a
subset of the axes therefore reads past the end of those vectors, and a
permuted axes list silently pads the wrong axis:

auto x = reshape(arange(25.0f), {5, 5});
pad(x, {1, 0}, Shape{1, 2}, Shape{1, 2}, array(0.0f), "edge");
// input lands on the wrong axis and 3 rows are left zero; no error

A debug build traps in SmallVector::operator[] with edge_pad at
ops.cpp:1551 on the stack. ASan is quiet at small ndim because Shape is a
SmallVector with 10 inline slots, so the read stays inside the object. Once
ndim puts the pad sizes on the heap it reports the overflow directly:

ERROR: AddressSanitizer: heap-buffer-overflow
READ of size 4 ... 0 bytes after 128-byte region
    #0 mlx::core::edge_pad(...)

reflect_pad indexes by position and is correct for a subset, but it does not
normalize a negative axis (ops.cpp:1468, 1476), so axes = {-1} indexes
starts[-1]:

small_vector.h:316: runtime error: addition of unsigned offset to
0x...060 overflowed to 0x...058

Axis handling across the pad family now goes through normalize_axis_index,
the same consolidation #4288 applied to split, unstack, partition and
topk. That helper also range checks, which changes behaviour in one place:
pad previously indexed out_shape with an unvalidated axis (ops.cpp:1621),
so an out-of-range axis was undefined behaviour for every mode. It now raises
std::invalid_argument. That is a new error path, not just a bug fix.

pad validates and normalizes every axis before the mode dispatch, so the
calls inside edge_pad and reflect_pad are redundant by construction.
Normalizing once in pad and passing the normalized vector down measures
+27/-20 on this file against +22/-18, and it changes what the Pad primitive
stores, and with it is_equivalent, state() and the axes Pad::vmap sees.
Keeping the helper at each site leaves the primitive untouched, but the
single-call version is a small change if you prefer it.

edge_pad landed in 635ccd9 (2024-08-06, #1309). reflect_pad is newer,
9f35f77 (2026-08-11, #3608), and already indexes by position; this adds the
axis normalization it was missing and brings edge_pad in line with it.
constant was already correct in both respects.

This is reachable from mlx_pad in mlx-c (mlx/c/ops.h:730), which forwards
caller-supplied axes, independent pad-size lengths, and a mode string. No
Python path reaches it: the bindings only build full-length axes
(python/src/ops.cpp:3493), which is also why the existing tests miss it, the
only C++ test on this overload using a 1-D array where axes.size() == ndim.

Checklist

Put an x in the boxes that apply.

• I have read the CONTRIBUTING document
• I have run pre-commit run --all-files to format my code / installed pre-commit prior to committing changes
• I have added tests that prove my fix is effective or that my feature works
• I have updated the necessary documentation (if needed)