[wasm] add flags for WebGPU build to reduce binary size

[fs-eire]

Description

This change uses a combination of flags to reduce binary size for WebGPU wasm build:

• exclude a few data types: uint16/int16, float8, float4, float64, sparse tensor, string
• disable_ml_ops
• disable_generation_ops

Motivation and Context

reduce binary size for ORT-web.

[fs-eire]

Updated the change to use config file and --enable_reduced_operator_type_support.

Pending more local test, now switching to "WIP" and will switch back once it's complete.

[xenova]

Very aligned with this fix, but we do get the following error for https://huggingface.co/onnx-community/BEN2-ONNX/tree/main/onnx.

ERROR_MESSAGE: /mnt/vss/_work/1/s/onnxruntime/core/optimizer/transformer_memcpy.cc:254 bool onnxruntime::TransformerMemcpyImpl::IsNodeCompatibleWithProvider(const onnxruntime::Node &) const !node_provider_type.empty() was false. Provider type for Pow node with name '/dec_blk3/norm1/Pow' is not set.

which uses float64 ops (first Pow).

This is VERY rare though... so probably not worth reverting/updating. We should update the model instead imo.

[fs-eire]

Very aligned with this fix, but we do get the following error for https://huggingface.co/onnx-community/BEN2-ONNX/tree/main/onnx.

ERROR_MESSAGE: /mnt/vss/_work/1/s/onnxruntime/core/optimizer/transformer_memcpy.cc:254 bool onnxruntime::TransformerMemcpyImpl::IsNodeCompatibleWithProvider(const onnxruntime::Node &) const !node_provider_type.empty() was false. Provider type for Pow node with name '/dec_blk3/norm1/Pow' is not set.

which uses float64 ops (first Pow).

This is VERY rare though... so probably not worth reverting/updating. We should update the model instead imo.

Yes usually fp32 works for most of the scenarios.