fix(cpu): correct NEON convolution padding and parameter names

Fix padding handling in NEON SIMD convolution kernels to properly handle
signed offsets and bounds checking. Update parameter names for consistency
(batch_size → batch, input_h/w → height/width, input_length → length).

Padding is now correctly applied by computing signed offsets and skipping
iterations that fall outside input bounds, matching the expected behavior
for padded convolutions.

Author: farhan-syah