cpu: some improvements

[dzarukin]

This is some groundwork for future asynchronous runtime support:

• There are overflow issues caught by -ftrapv option when running test_conv_all for conv and ip.
• There's a rename for the variable which is shadowed by outer same variable. It get a problem when capture by copy is used.
• Make post-ops kernel apply scales when they are defined in attributes. It's possible to distinguish them since v3.0.
• Align scale_utils function interfaces to rely on same objects between calls for proper re-use of the code underneath.

(there's more to come).

[dzarukin]

make test
disable test_device_gpu

[dzarukin]

make test
disable benchdnn_all
enable benchdnn_conv
disable test_device_gpu
enable benchdnn_nightly