DQLinear: Add support for per_token activation and fp16 dtype (#2327)

Summary:
Pull Request resolved: https://github.com/pytorch/executorch/pull/2327

Wire per_token quantization support in xnnpack delegate. ATM PT2e quant flow (XNNPACKQuantizer --> prepare --> convert) doesn't support `q/dq` nodes for `per_token`. So added eager mode hacks to generate such graph in tests. Once we add support the graph should remain the same so it is ok to test lowering logic like that.

bypass-github-export-checks

Reviewed By: kimishpatel, mcr229

Differential Revision: D54026295

fbshipit-source-id: 269559d321674b21f18352c7e1ed649665963567