Define min_dot_size outside of XPUBackend

[whitneywhtsang]

min_dot_size was moved inside XPUBackend in #5329, which makes user experience inconsistent with other backends.

[whitneywhtsang]

What do reviewers think of moving min_dot_size back out?

[chengjunlu]

What do reviewers think of moving min_dot_size back out?

It has to be removed back out for supporting the Helion.

[anmyachev]

What do reviewers think of moving min_dot_size back out?

It's fine if we don't break anything. @AndreyPavlenko please take a look.

[AndreyPavlenko]

It has been moved because it's backend specific. Different backends may have different implementations.
What's the issue of using it as a class method? Here is an example - #5329 (comment)

[whitneywhtsang]

It has been moved because it's backend specific. Different backends may have different implementations. What's the issue of using it as a class method?

Different from CUDA, which make users like Helion to have xpu specific changes.

[AndreyPavlenko]

What about something like this?

def min_dot_size(device_props: dict):
    backend = XPUBackend(triton.runtime.driver.active.get_current_target())
    return backend.min_dot_size(device_props)

[whitneywhtsang]

What about something like this?

def min_dot_size(device_props: dict):
    backend = XPUBackend(triton.runtime.driver.active.get_current_target())
    return backend.min_dot_size(device_props)

It should work, changed.