DirectML EP: add Tier-2 op+activation fusion, fp16 exec flags, and ORT parity fixes

[KenLagos]

Add a generic Tier-2 BaseOp+Activation fusion rule and align several DirectML plugin behaviors with ONNX Runtime's DirectML execution provider.

Op+Activation fusion (new):

• Add op_activation_ep_fusion.{h,cc} and fusion_activation_helper.h, fusing an optional downstream activation into Add, Sum, Gemm, Conv, ConvTranspose, BatchNormalization, InstanceNormalization, and MeanVarianceNormalization via the DML operator's FusedActivation field, with allowed activations filtered per op and per device to match ORT's fusability tables.
• Fix a bug where the eager compile path never recorded compiled_slot_sizes, causing the fused kernel to be needlessly recompiled on every inference instead of once (~2x perf regression on fused Conv/ConvTranspose models).

Runtime (non-constant) weight support:

• Bind Conv/ConvTranspose/BatchNorm/InstanceNorm weight, scale, bias, mean, and variance inputs per slot, either directly from a runtime OrtValue or via a one-time static upload for true constants, and defer kernel compilation until all weight shapes are known.

ORT parity fixes:

• BatchNormalization epsilon now defaults to 0.0 (InstanceNormalization keeps 1e-5), matching ORT's per-op defaults.
• MeanVarianceNormalization explicit axes are normalized for negative values, matching ORT's GetDmlAdjustedAxis.

fp16 execution flags:

• Set DML_EXECUTION_FLAG_ALLOW_HALF_PRECISION_COMPUTATION when a graph is entirely fp16, in full-graph/per-partition compilation and in the QuickGelu, BiasGelu, and FusedMatMul fusion kernels.

Added a follow up

Fix Tier-0 graph fusion binding validation failure for ConvInteger models
The owned_graph_input_indices construction used raw packed DML slot
indices into the ONNX-ordered input_names array, causing initializer
weights to be looked up under wrong names for operators with
input_name_reorder (e.g. ConvInteger). This left OWNED_BY_DML
initializers unbound in the init binding table, triggering a DirectML
BindingValidator E_INVALIDARG and device removal during session init.

Additional fixes:

• Resolve passthrough aliases in consumed_initializer_names tracking
  so initializers consumed through renamed edges get DML graph input slots
• Use min(actualTensorBytes, totalTensorSizeInBytes) for constant node
  DataSize to avoid reading past ORT tensor allocations
• Add dml_input_slot_indices to ConvInteger, QLinearConv, MatMulInteger,
  QLinearMatMul, and MatMulIntegerToFloat for correct edge wiring when
  optional inputs are absent
• Fix filter zero-point and scale axis placement from 0 to 1 (C-axis)
  in ConvInteger and QLinearConv per DML spec
• Guard InitializeOperator/Flush with ORT_TRY/ORT_CATCH so a single
  bad partition falls back instead of aborting the process
• Extract DrainDmlDebugMessages helper and add device-removed HRESULT
  logging in CloseAndExecute
• Rename ambiguous variables (cn, sg, im, npi, ni, ib, ci) for
  readability