[CPU] Eliminate cross-dispatch pack/unpack overhead when tensor.pad intervenes between producer and consumer #24914

Request description

When a tensor.pad operation sits between a producer dispatch's output and a consumer dispatch's encoded input (e.g., convolution activations with spatial padding), the current pipeline creates a cross-dispatch boundary sequence of unpack → pad → repack. This three-operation overhead is avoidable when the padding is constant and symmetric, since the CPU can apply the padding implicitly.

Current behavior:
producer_dispatch (tiled output)
→ unset_encoding (unpack to standard layout)
→ tensor.pad (add spatial halos)
→ set_encoding (repack to tiled layout)
→ consumer_dispatch (tiled input)

Desired behavior:
producer_dispatch (tiled output, unpadded shape)
→ consumer_dispatch (reads tiled unpadded buffer)

Proposed approach:

1. In SetEncoding, detect symmetric constant tensor.pad on convolution activations and encode the unpadded source instead
2. Strip kernel-offset reduction addends from the convolution's activation indexing maps to match the smaller operand extent
3. Make tensor.pad hoistable and fusable with set_encoding in HoistEncodingOps and FuseEncodingOpsIntoDispatchRegions
4. Add tensor.pad to the "look-through" chain in FusionUtils::getProducerDispatchValueAndOpChain()
5. Store padding metadata as attributes for the backend to program hardware padding

Eliminates per-layer unpack/repack overhead in CNN inference pipelines where every convolution layer pads its activation input.

What component(s) does this issue relate to?

Compiler

Additional context

No response