[quant][fx] Don't assume bias is a keyword argument (#61647)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/61647
`prepare_fx` currently assumes that bias is always a positional argument to
convolutions, and only a keyword argument to other functions. This happens to work
today due to a quirk in how `__torch_function__` is handled for python
functions but shouldn't be considered stable.
Instead, we should support `bias` for both positional and keyword forms.
cc jerryzh168 jianyuh raghuramank100 jamesr66a vkuzo
Test Plan: Imported from OSS
Reviewed By: ngimel
Differential Revision: D31401360
Pulled By: albanD
fbshipit-source-id: 1e2f53d80e2176b870f326dc498e251e2386136e
