Reserve WebGPU program vector capacity

[Lingavasan]

Fixes #28516

Summary

• add ProgramBase reserve helpers for inputs, outputs, and uniform variables
• reserve capacity when bulk-adding inputs, outputs, and uniform variables
• apply explicit capacity hints in WebGPU convolution program setup paths
• add compile-time checks that the new reserve helpers preserve chain-style ProgramBase& returns

Validation

• git diff --check
• PATH="$PWD/.venv/bin:$PATH" .venv/bin/lintrunner -a onnxruntime/core/providers/webgpu/program.h onnxruntime/core/providers/webgpu/program.cc onnxruntime/core/providers/webgpu/program_test.cc onnxruntime/core/providers/webgpu/nn/conv.cc onnxruntime/core/providers/webgpu/nn/conv2d_mm.cc
• PATH="$PWD/.venv/bin:$PATH" python3 tools/ci_build/build.py --build_dir /private/tmp/ort-webgpu-build --config Release --update --use_webgpu --cmake_path "$PWD/.venv/bin/cmake" --cmake_generator Ninja --skip_tests
• PATH="$PWD/.venv/bin:$PATH" .venv/bin/ninja -C /private/tmp/ort-webgpu-build/Release onnxruntime_providers_webgpu
• PATH="$PWD/.venv/bin:$PATH" .venv/bin/ninja -C /private/tmp/ort-webgpu-build/Release onnxruntime_test_all onnxruntime_provider_test onnx_test_runner
• PATH="$PWD/.venv/bin:$PATH" .venv/bin/ctest --test-dir /private/tmp/ort-webgpu-build/Release --output-on-failure

CTest note: onnx_test_pytorch_converted and onnx_test_pytorch_operator passed locally. onnxruntime_test_all and onnxruntime_provider_test built successfully but abort on this local macOS environment because Dawn reports Failed to get a WebGPU adapter: No supported adapters while initializing the WebGPU EP, before completing the suite.

[Lingavasan]

@microsoft-github-policy-service agree