Fix cache collisions and zero-initialize decompression scratchpad in InnerProduct

[Shanmuk4622]

This PR resolves non-deterministic NaN outputs observed in the OpenVINO CPU plugin (OpenVINO issue #36328) during weight decompression (specifically decompressing f8e8m0 weights to f32).

Problem:

1. Cache Key Collision: Custom OpenVINO weight decompression parameters were missing from quant_entry_t::get_hash(), serialize(), and deserialize(). This led to hash collisions and incorrect weights sharing across different compiled models.
2. Uninitialized Decompression Scratchpad Buffer: The decomp_buf_global memory fetched from the oneDNN scratchpad (key_brgemm_primitive_decomp_buf) was not zero-initialized. When the subsequent GEMM executor performed aligned vector loads, it read uninitialized padding elements containing garbage/NaN bits. Multiplying them by 0 (since they were out-of-bounds/padded) still produced NaN outputs (0.0 * NaN = NaN), propagating NaNs non-deterministically.

Solution:

• Added OpenVINO scaling and zero-point decompression parameters (type, mask, data type, dims) to quant_entry_t::get_hash(), serialize(), and deserialize().
• Zero-initialized the allocated scratchpad decompression buffer decomp_buf_global using std::memset before executing decompression in jit_brgemm_inner_product.cpp.

Fixes: Related to openvinotoolkit/openvino#36328

Checklist

General

• Do all unit and benchdnn tests pass locally for each commit? (Verified via local compilation and custom reproduction script)
• Have you formatted the code using clang-format?

Bug fixes

• Have you included information on how to reproduce the issue? (Yes, documented in OpenVINO issue #36328)
• Have you added relevant regression tests?

[Shanmuk4622]

Technical Review: oneDNN Decompression Submodule Fixes

1. Hashing, Serialization & Deserialization (primitive_attr_quant.cpp)

• Correctness: The additions of type_, is_set_, is_set_scale, mask_scale, data_type_scale, ndims_scale, dims_scale, is_set_wei, mask_wei, data_type_wei, ndims_wei, and dims_wei to get_hash(), serialize(), and deserialize() are symmetric and maintain exact parity.
• Style & Conventions: Using hash_combine for scalar values and primitive_hashing::get_array_hash for the shape arrays is correct and aligns with oneDNN's existing hashing conventions.
• Impact: This cleanly eliminates weights cache collisions between models with distinct decompression parameters.

2. Scratchpad Zero-Initialization (jit_brgemm_inner_product.cpp)

• Design & Timing: Performing the std::memset call in the master thread immediately after retrieving decomp_buf_global is optimal. It avoids redundant or multi-threaded initializations inside the parallel loop.

• Memory Safety: The size computation logic:

  • For weights_compressed (using jbgp.ic * 64)
  • For weights_decompression in prepack mode (using jbgp.ic_block * jbgp.nb_ic_blocking * jbgp.oc_block)

  exactly matches the sizes allocated during configuration booking (init_scratchpad_base). This makes the std::memset boundary completely safe against buffer overruns.

• Impact: It successfully forces all alignment and boundary elements to 0.0f, preventing vector instructions from loading uninitialized garbage and propagating NaNs.