[GPU] Fall back to ocl for convolution with unaligned feature

[davidsnam-intel]

Issue

• A dynamic-shape model produces NaN at a Convolution whose input/output feature count is not a multiple of the blocked-format sub-feature size.

Root Cause

• For an unaligned feature (85), the blocked format b_fs_yx_fsv16 pads the feature dimension up to a multiple of 16 (96). The oneDNN convolution primitive reads the full sub-feature blocks, including the padding lanes.
• In the dynamic memory-reuse path the reused buffer is not zero-filled, so those padding lanes contain stale bytes.
• The concatenation oneDNN impl already guards against exactly this contract (zero-filled fsv padding required for unaligned feature blocks), but convolution did not.

Solution

• Add an is_feature_aligned() check to the oneDNN convolution validate_impl()
• If the input or output uses a blocked format whose feature dimension is not divisible by the sub-feature block size, reject the oneDNN impl so the node falls back to the OCL convolution, which handles the padding lanes safely.

Tickets:

• CVS-190448