Account for additional local memory requirements in Global dispatch

[hjabird]

• Not accounting for all local memory requirements in dispatch causes errors on A100.
• This PR accounts for the additional required memory.
• And also adds a regression test.
• This is not an ideal fix for the issue - the memory requirements depend on the ordering of the sub-impls. This information is not available at dispatch time. A more significant refactor is required.

Checklist

Tick if relevant:

• [n/a] New files have a copyright
• [n/a] New headers have an include guards
• [n/a] API is documented with Doxygen
• New functionalities are tested
• Tests pass locally
• Files are clang-formatted

[hjabird]

@AD2605 is it reasonable to summarise your changes as adjusting the global implementation to adjust the number of sub-groups used in a kernel to that set by num_scalars_in_local_mem instead of using PORTFFT_SGS_IN_WG?

[AD2605]

@AD2605 is it reasonable to summarise your changes as adjusting the global implementation to adjust the number of sub-groups used in a kernel to that set by num_scalars_in_local_mem instead of using PORTFFT_SGS_IN_WG?

Yes, that would be correct and was the cause of the error

[hjabird]

As commit author, the Github interface doesn't let me approve what has become Atharva's PR. But LGTM.

[AD2605]

Given that I have an approval from Hugh, Approving this from my side to hit the 2 approval criteria and going ahead wuth the merge.