fix: handle non-multiple-of-4 inputs in histogram_i32x4_kernel

[Ajith-Kumar-Nelliparthi]

Fix histogram_i32x4_kernel boundary handling

Problem

The current implementation loads 4 integers unconditionally using INT4(a[idx]).

When N is not divisible by 4, the kernel may read out of bounds because only idx < N is checked.

Fix

Added a safe fast path for full int4 loads using:

if ((idx + 3) < N)
Added tail handling for remaining 1-3 elements.
Result

This prevents out-of-bounds memory accesses for non-multiple-of-4 input sizes while preserving the vectorized fast path.