Fix Upsample align_corners=False for fractional scale factors

[v-code01]

Problem

_scaled_indices computes the align_corners=False start offset as

start = ((M - 1) * step - N + 1) / 2   # M = int(scale * N), step = 1 / scale

This equals the half-pixel offset (1 - step) / 2 only when M == scale * N exactly. For a fractional scale factor M = int(scale * N) is floored, so the grid is re-centered to the corner-aligned (align_corners=True) grid. The result: align_corners=False output becomes bit-identical to align_corners=True, i.e. the flag is silently ignored.

x = mx.arange(5).reshape(1, 5, 1).astype(mx.float32)
f = nn.Upsample(scale_factor=1.5, mode="linear", align_corners=False)(x)
t = nn.Upsample(scale_factor=1.5, mode="linear", align_corners=True)(x)
# f == t == [0, 0.667, 1.333, 2.0, 2.667, 3.333, 4.0]
# correct half-pixel: [0, 0.5, 1.167, 1.833, 2.5, 3.167, 3.833]

This affects linear and cubic modes (both route through _scaled_indices) at any fractional scale (1.5, 1.25, ...). All existing upsample tests use integer scale factors, where the two expressions coincide, so the regime was untested.

Fix

Use the constant half-pixel offset start = (1 - step) / 2, so the sampling positions are (i + 0.5) / scale - 0.5 for every scale. Integer scale factors are unchanged.

Test

Added test_upsample_align_corners_false_fractional_scale pinning the half-pixel positions at scale=1.5 and asserting align_corners=False now differs from True. It fails on the old code and passes with the fix; the existing upsample tests stay green.