benchdnn: improve memory estimation

[dzarukin]

This PR brings a bunch of general improvements:

• Add "run ref" status line so that it's easier to distinguish between GPU and CPU problems when validating GPU cases with CPU reference.
• Use mprotect for --mode=R, previously it was used for --mode=C only. This brings a consistent behavior to running the test case with the library side.
• The major one - improves memory estimation for prim_ref objects. It usually happens that CPU ref primitive requires weights in blocked format and padding is possible in this case. In real extreme cases, it could lead up to 64x more memory for weights that it was estimated for reference because original estimation would consider ref memories in f32 abx format. With this PR, the logic would collect real sizes of prim_ref, combine them with the library primitive and perform a check. If it doesn't fit, current prim_ref wouldn't be used (it's allowed up to 4 prim_ref versions). If none of them fits but the library one fits with pure ref, that one will be executed.

[dzarukin]

make test
set test_scope=NIGHTLY

[dzarukin]

make test
set test_scope=NIGHTLY