[GPU] Fix oneDNN FP16 convolution format selection for channel expansion operations

[andrew-k-park]

Details:

• When FP16 dynamic convolution has small input channels (≤4) and large output channels (e.g., 1024), the current format selection logic chooses bfyx → fsv16, which triggers oneDNN reference kernel instead of optimized JIT kernel, resulting in significant performance degradation.
• Override output format to planar (bfyx) when input channels are small (≤ 16), and output channels are large (≥ 32)

Current behavior:

• Input: 3 channels → Converted to bfyx
• Output: 1024 channels → Remains fsv16 (only changed when output ≤ 4)
• Result: bfyx → fsv16 combination uses reference kernel (slow)

Root Cause

The fsv16 blocked format is optimized for reading many channels but introduces overhead when used for writing outputs in channel-expansion scenarios (small input → large output). oneDNN's reference kernel is selected because:

1. Inefficient write pattern: fsv16 output requires interleaved writes every 16 elements (non-contiguous)
2. No optimized implementation: oneDNN doesn't provide JIT-optimized kernel for fsv16 output generation from small input channels
3. Scatter write overhead: Writing 1024 channels in fsv16 format requires complex block-strided access

Tickets:

• CVS-177671