[GPU][NVL-P] Use upconversion for unsupported scales on NVL-P

[kealan-barbieri]

Description

Fix NVFP4 support on NVL-P, use upconversion to avoid group size limitations on late scaling that conflict with NVFP gs16.

Fixes # MFDNN-14876

EDIT: added fix for last remaining layer failing with OOR in jit:

--matmul --engine=gpu --allow-enum-tags-only=false --check-ref-impl=true --stag=acb --wtag=acb --dtag=abc --attr-fpmath=tf32 16x64x64:16x64x15000_npointnet.tr.tf32.pt.mb16*1

Checklist

General

• Do all unit and benchdnn tests (make test and make test_benchdnn_*) pass locally for each commit?
• Have you formatted the code using clang-format?

Bug fixes

• Have you included information on how to reproduce the issue (either in a github issue or in this PR)?
• Have you added relevant regression tests?

[kealan-barbieri]

make test
set test_scope=NIGHTLY
disable test_device_cpu
disable benchdnn_all
enable benchdnn_matmul
enable benchdnn_ip
enable arch_gpu_xe3p-lpg

[kealan-barbieri]

make test
set test_scope=NIGHTLY
disable test_device_cpu
disable benchdnn_all
enable benchdnn_matmul
enable benchdnn_ip
enable arch_gpu_xe3p-lpg