[MPS] Fix sum reduction saturated cast for integer overflow

[wdvr]

Fixes pytorch#179415

Problem

MPS backend's sum reduction uses [MPSGraph castTensor:toType:] to convert the float32 reduction result back to the original integer dtype. This API performs saturated casting (clamps to type range), but CPU uses wrapping (preserves only least significant bits).

Example:

torch.tensor([[-1, 2, 1], [5, 3, 6]], dtype=torch.uint8, device="mps").sum(dtype=torch.uint8)
# MPS (before fix): 255 (saturated)
# CPU (expected):   16  (wrapping: 272 % 256)

Fix

Before the final castTensor call in reduction_out_mps(), apply floor-modulo wrapping for small integer output types (uint8, int8, int16):

• Unsigned: value - floor(value / range) * range (equivalent to value % range)
• Signed: shift to unsigned range, modulo, shift back

This ensures the float32 value is already within the valid type range before the saturated cast, making both casts equivalent.

Test plan

• Added regression test test_sum_integer_overflow_wrapping in test/test_mps.py
• Tests uint8 overflow: [255, 2, 1, 5, 3, 6].sum() == 16
• Tests int8 overflow: [127, 1].sum() == -128