Support exporting QNN models with Python wheels out-of-the-box

[jathu]

As part of #9019, we recently added CoreML export support "out of the box" for the executorch pip package: #9483. Developers can now export to CoreML backend, by just pip installing executorch — previously this required cloning the ExecuTorch repo, and building everything from source. We anticipate this to encourage ET adopting for CoreML.

We want to provide the same developer experience when exporting models to Android on Qualcomm chips.

Currently, this requires developers to manually download the QNN SDK, clone the ExecuTorch repo, and build everything from source. Similar to CoreML, we should bundle this requirement into ExecuTorch. Given the QNN library is currently only distributed as a zip file, here are some potential next steps:

• [Option 1] [Preferred] Ideally, Qualcomm should provide the SDK through pip so we can just include that as a dependency. This is what we currently do with CoreML
• [Option 2] [Preferred] While we wait for Qualcomm, we can temporarily bundle the library into executorch. This will require permission from Qualcomm
• [Option 3] [Not ideal] At the very least, we should automate downloading the SDK in the build script

cc @larryliu0820 @cccclai @winskuo-quic @shewu-quic @cbilgin @lucylq

[cccclai]

cc: @shewu-quic @shewu-quic @haowhsu-quic @winskuo-quic @DannyYuyang-quic

Currently we require users to download the qnn library and a script to install, we may need to figure out

1. can the qnn library downloading part of the script
2. can qnn library be part of the wheel package