[GPU][TESTS] Fix weightless caching tests

[tkrupa-intel]

Currently the tests have weightless caching enabled, but the models tested contain no constants eligible for weightless caching. This fixes the situation.

To elaborate - all the models have f32 dtype constants by default. In these particular models all of the constants which would otherwise be eligible for weightless caching undergo some precision transformation which invalidates this eligibilty. The PR changes the dtype of the constants to prevent that and deletes the models which contain no constants eligible for weightless caching even after the dtype change.

[p-durandin]

build_jenkins

[e-ddykim]

Minor comments:
The root cause is that the default inference precision of GPU is f16.
You can set the inference precision as below:
ov::AnyMap config = { ov::hint::inference_precision(ov::element::f32), ... };
Looks good to me

[p-durandin]

to be rebased after #27171 merge