Implement BatchnormBwd functionality into hip-kernel-provider

[BalintCsala]

Caution

Must be merged after #4885

Motivation

This PR moves the BatchnormBwd implementation of MIOpen into the hip-kernel-provider hipDNN plugin.

Technical Details

• Backward batchnorm plan (BatchnormBwdPlan): Implements miopenBatchNormalizationBackward_V2 with support for both spatial and per-activation modes. Selects between single-kernel and multi-kernel (stash-based) variants via a heuristic based on tensor dimensions.
• Fused backward+activation: Supports miopenBatchNormBackwardActivation by fusing a pointwise activation backward pass (ReLU, sigmoid, tanh, ELU, etc.) into the batchnorm backward kernel.
• Plan routing (BatchnormPlanBuilder): Extended isApplicable() and buildPlan() to handle single-node backward (1-node graph) and fused backward+activation (3-node graph: BnInference → Pointwise → BnBackward).
• Applicability checks: Added checkBatchnormBackwardTensorConfigSupported(), checkBatchnormInferenceActivationBackwardTensorConfigSupported(), and checkBatchnormBwdActivationModeSupported() for validating tensor layouts and data types.
• Activation utilities (HipKernelUtils): Added ActivationMode enum, ActivationParams struct, and parseActivation() to map PointwiseAttributes to activation parameters used by backward kernels.
• HIP kernels: Added BatchNormBwdSpatial.cpp, BatchNormBwdSpatialMultiple.cpp (multi-kernel stash variant), BatchNormBwdPerAct.cpp, and BatchnormStash.hpp.
• Integration tests: Backward and backward+activation tests across NCHW/NHWC/NCDHW/NDHWC layouts with fp32/fp16/bfloat16.
• Unit test: Validates BatchnormBwdParams construction for single backward and fused activation cases.

Test Plan

Testing done through included unit and integration tests.

Test Result

Tests passed on gfx90a

Submission Checklist

• [ X] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.