Handle untyped interfaces in _check

[maxmynter]

Guard match checks of types in function signature and supplied via the params kwarg in the benchmark decorators _check method against untyped function signatures.

If untyped, we do not check for matching types. We also log a warning to the stdout of the logger indicating the untyped parameter and function name.

[nicholasjng]

With this change, the following should be a failure (non-unique types):

import nnbench

@nnbench.benchmark
def increment(value):
    return value + 1


@nnbench.benchmark
def double(value: int) -> int:
    return value * 2

Is this expected? Are we fine with this?

(We should check that the resulting type in the error message is still sensible, though. Is it <class 'empty'> by any chance?)

[nicholasjng]

Suggested change

	if not param.annotation == inspect.Parameter.empty and not issubclass(
	param_types[name], param.annotation
	): # only check type match if type supplied
	# skip the check if benchmark interface is untyped
	if param.annotation == inspect.Parameter.empty:
	continue
	if not issubclass(param_types[name], param.annotation):

[maxmynter]

(We should check that the resulting type in the error message is still sensible, though. Is it <class 'empty'> by any chance?)

It is <class 'inspect._empty'>.

[nicholasjng]

(We should check that the resulting type in the error message is still sensible, though. Is it <class 'empty'> by any chance?)

It is <class 'inspect._empty'>.

It's fine for now. We can think about making it show up as <untyped> or something, but that requires some more effort, and would only be relevant in the case of an error, i.e. just before raising the exception.

[maxmynter]

i would argue that your example above should throw an error. Untyped and typed might very well be non-unique types. We cannot handle different types for the same variable name and if untyped cannot ensure it is the right one. So it's either on the user to keep everything untyped and ensure the right value is passed. Or the absence of the type is an oversight in which case the error is a welcome reminder.