Divide as a * (1.0 / b) during weight compression

[nikita-savelyevv]

Changes

• Introduce fns.reciprocal(x) = 1.0 / x operation
• Replace regular division a / b with a * fns.reciprocal(b) during weight compression -related computations.

Model	Develop PPL	Branch PPL	Develop Similarity	Branch Similarity
phi3-mini int4_asym	10.96023	10.95891 (-0.012%)	0.89861	0.89635 (-0.25%)
phi3-mini int4_asym SE	10.59984	10.60182 (+0.019%)	0.88692	0.88924 (+0.25%)
phi3-mini int4_asym GPTQ	10.73838	10.72351 (-0.139%)	0.88555	0.86824 (-1.96%)
phi3-mini int4_asym GPTQ + SE	10.51594	10.50970 (-0.059%)	0.871897	0.89654 (+2.82%)

Reason for changes

Align with how division is performed in OpenVINO. Prerequisite to #2727. This should be merged first in order to later track only the differences introduced by the OV computations.

Related tickets

139047

Tests

Added a test comparing inverted division result to a reference value.
NNCF/job/manual/job/post_training_weight_compression/271/
NNCF/job/manual/job/post_training_quantization/576/

Also weekly validation is requested in 158012

[nikita-savelyevv]

NNCF/job/manual/job/post_training_weight_compression/267/ has passed.

[alexsu52]

LGTM

[nikita-savelyevv]

The problem with division is resolved by aligning OV division with NP division by disabling ConvertDivide and ConvertDivideWithConstant transformations. These transformation can be disabled for target division nodes by setting "nonconvertable_divide_0" rt_info flag.