This reverts commit 2804e181.

'inputs_like' was a unnecessary argument to make the implementation
easier. But that led to a confusing interface. Now the compiled operator
is smart enough to do the inference of the input data types on its own.
It maintains a mapping from the input dtypes it saw to the executable.

Went a step ahead and clarified that such layout-ordering assumptions
would be taken throughout PytatoArrayContext.

* mimic numpy's norm when ord==None

- When ord==None, numpy ravels the array and then computes the euclidean
  norm.

* test norm when ord is None

* fixup! mimic numpy's norm when ord==None

Co-authored-by: Andreas Klöckner <inform@tiker.net>

* implement PyOpenCLArrayContext.ravel

* fixup! implement PyOpenCLArrayContext.ravel

- order="K" was only correct for >0 stride values, instead renamed it to
  order="A" as per numpy.

Co-authored-by: Andreas Klöckner <inform@tiker.net>