[GPU] Fix OOB reads in convolution bfyx_os_iyx_osv16 for multi-batch

[davidsnam-intel]

Issue

• DepthAnythingV2 inference fails with CL_OUT_OF_RESOURCES on Xe2+ when running with batch ≥ 2.
• Pre-Xe2 silently returns zero for out-of-bounds reads, so the bug was latent on older platforms.
• But, Xe2+ raises CL_OUT_OF_RESOURCES.

Root cause

• With batch ≥ 2, gws includes dummy output tiles beyond the valid output region for the last batch.
• These tiles compute input addresses that exceed the tensor's physical length
  (INPUT0_OFFSET_WITH_PADDING + INPUT0_BATCH_PITCH * INPUT0_BATCH_NUM)

Solution

• Add [min(idx, input0_physical_len - 1)] for index clipping.
• This is the pattern already used by every other input-read path in the same kernel.
• The clamped value is never used in actual output computation since dummy tiles are discarded by the output-boundary guard.
• So, this change has no effect on inference accuracy or performance — it only prevents the hardware fault from the out-of-bounds memory access.

Tickets

• CVS-186560